A context-aware and human-centered framework for road safety assessment using semantic scene understanding

Abstract

Purpose:

Assessing perceived road safety from street imagery remains a challenge because human judgments depend on infrastructure, context, and subjective interpretation. Current computer vision methods focus on object detection and semantic segmentation, with limited attention given to human-centric safety. This study proposes a Context-Aware Road Safety Assessment (CARSA) framework based on semantic scene understanding and multidisciplinary expert evaluation.

Methods:

A semantically guided image selection pipeline was applied to the Mapillary Vistas dataset to construct a representative set of road scenes. Following manual filtering, the images were classified into urban, residential, and rural contexts. A literature-informed, rule-based Context-Aware Road Safety Assessment (CARSA) framework was then defined using semantic scene features, predefined context-specific factor weights, and categorical decision thresholds. All framework parameters were fixed and CARSA predictions were generated before the expert annotation process. The images were subsequently evaluated independently by experts in transportation engineering, driving instruction, and computer vision, and the resulting expert consensus was used exclusively as an independent Ground Truth for paired human–framework agreement evaluation on the current benchmark.

Results:

The expert evaluation demonstrated excellent reliability, achieving an ICC(A,k) of 0.953, a Cronbach’s Alpha of 0.954, and a Fleiss’ Kappa of 0.805. CARSA achieved an accuracy of 92.38%, a weighted F1 score of 92.61%, a Cohen’s Kappa of 0.853, and a Quadratic Weighted Kappa of 0.891. The non-contextual semantic baseline achieved 75.87% accuracy, whereas context-aware CLIP zero-shot, Qwen2.5-VL zero-shot, and Qwen2.5-VL few-shot achieved 19.68%, 20.32%, and 29.84%, respectively. CARSA significantly outperformed all four baselines in paired exact McNemar tests after Holm correction. Most CARSA disagreements were limited to adjacent safety categories.

Conclusion:

The proposed framework provides a reliable and explainable approach for assessing perceived road safety by integrating semantic scene understanding with contextual reasoning. Furthermore, the open benchmark, expert annotations, and evaluation protocol establish a reproducible foundation for future research in interpretable transportation systems, human–AI alignment analysis, and road scene understanding.

1 Introduction

Understanding perceived road safety from street-level imagery is an important challenge in Intelligent Transportation Systems (ITS) and urban analytics. Road scenes involve complex interactions among infrastructure, vehicles, pedestrians, and cyclists, and their safety implications cannot be determined through visual detection alone. Effective assessment therefore requires both accurate scene understanding and contextual interpretation (Dolatyabi et al., 2025; Guo et al., 2024).

Computer vision has substantially advanced transportation perception through object detection, semantic segmentation, lane detection, traffic-sign recognition, and transformer-based scene analysis (Muhammad et al., 2022; Ankareddy and Delhibabu, 2025). Related computer-vision systems have also demonstrated value in supporting human interpretation and decision-making in other application domains (Tbaileh and Bagriyanik, 2025). Street-level datasets such as Mapillary Vistas provide dense pixel-level annotations of roads, sidewalks, vehicles, pedestrians, signs, and surrounding infrastructure, enabling detailed analysis of diverse transportation environments (Hussain et al., 2022; Randhawa et al., 2025). Recent learning-based studies have further begun to model human perception directly from street imagery by combining semantic and visual representations (Huang et al., 2023).

Despite this progress, perceived road safety remains difficult to assess. Existing methods frequently emphasize the detection or segmentation of individual scene elements without fully representing how their safety meaning changes across environments (Yu et al., 2024; Kuznietsov et al., 2024). For example, the absence of a sidewalk may indicate inadequate infrastructure in an urban setting but may be expected on a rural road. Perceived safety therefore depends on how visible infrastructure is interpreted within its environmental context rather than on isolated objects alone.

Perceived road safety also differs from objective safety measures based on crash statistics or formal infrastructure audits. Whereas objective measures quantify recorded risks and compliance, perceived safety reflects how individuals interpret environmental and visual cues. Both perspectives can influence transportation behavior, urban planning, and infrastructure use (Hamim and Ukkusuri, 2024; Wang et al., 2024; Ye et al., 2024).

These characteristics motivate explainable and human-centered approaches that make the basis of safety judgments transparent and support trustworthy human–AI interaction (Graefe et al., 2023; Ridley, 2025). Within transportation and GeoAI, related work has further emphasized explainability, contextual adaptation, and human expertise (Yang et al., 2023; Ye et al., 2025a). Street-view research has shown that perceived safety is influenced by road configuration, built-environment characteristics, visual attention, and subjective interpretation (Cui et al., 2023; Liu et al., 2025). However, existing studies rarely combine semantic scene information, explicit road-context reasoning, and multidisciplinary expert evaluation within a single reproducible framework.

Assessing perceived road safety requires complementary professional perspectives. Transportation engineers emphasize infrastructure and regulation, driving experts consider visibility, behavior, and conflict exposure, and computer vision researchers focus on visual and semantic interpretation. Research nevertheless remains limited on how these perspectives can be integrated to evaluate semantically rich road scenes across Urban, Residential, and Rural environments. Mapillary Vistas provides detailed semantic annotations but does not include perceived-safety labels, and no widely available benchmark currently captures multidisciplinary expert judgments for this three-context task.

To address this gap, this study introduces CARSA, a literature-informed and rule-based framework for contextual perceived road-safety assessment. Images were selected through semantic-richness-guided sampling, filtered for human relevance, and categorized into Urban, Residential, and Rural contexts. CARSA then combined predefined semantic factors, context-specific weights, and interpretable decision thresholds. All framework components were finalized and all predictions were generated before expert annotation, as shown in Figure 1.

FIGURE 1

The retained images were independently evaluated by experts in transportation engineering, driving instruction, and computer vision using Low-, Medium-, and High-safety categories. The experts were blinded to CARSA outputs, and their consensus annotations were used exclusively as an independent reference standard for human–framework agreement. No rule, factor weight, threshold, or scoring component was selected, optimized, or recalibrated using the expert annotations. CARSA is intended to approximate perceived road-safety reasoning through transparent semantic and contextual rules rather than replace engineering audits or accident-risk models, as shown in Figure 2. The resulting benchmark also supports future research on vision–language models and human–AI agreement (Wu et al., 2026; Yin et al., 2026).

FIGURE 2

2 Related work

2.1 Computer vision for road scene understanding

Computer vision is a central component of Intelligent Transportation Systems, autonomous driving, and advanced driver-assistance systems, supporting the automated interpretation of complex road environments. Semantic segmentation and transformer-based architectures have substantially improved dense scene understanding and the modeling of spatial relationships, although contextual interpretation and practical deployment remain challenging (Muhammad et al., 2022; Ankareddy and Delhibabu, 2025).

Recent segmentation research has therefore focused on extracting richer contextual information from transportation scenes. Modified CANet architectures have demonstrated the value of multi-scale representations, while SUSC-SNet uses Swin-UNet and adaptive skip connections to improve the segmentation of small objects and complex road boundaries (Khairnar et al., 2025; Li J. et al., 2025). Large-scale datasets such as Mapillary Vistas support this development through geographically diverse pixel-level annotations, and Mapillary-based analyses of roadside infrastructure have shown the value of integrating multiple scene elements for ITS and high-definition map reconstruction (Hussain et al., 2022; Yang et al., 2022).

The field has also expanded from scene parsing toward broader environmental interpretation. Large-scale analyses of road-surface characteristics and semantic urban environments demonstrate that transportation scenes vary substantially across geographic and functional contexts (Randhawa et al., 2025; Yap and Biljecki, 2023). Reviews of traffic-scene understanding similarly identify environmental variability, dynamic conditions, contextual reasoning, and explainability as continuing challenges. Reasoning-oriented approaches such as Logic-RAG further illustrate that object recognition alone may be insufficient for context-aware transportation decisions (Dolatyabi et al., 2025; Kabir et al., 2025).

This shift has increased interest in transparent transportation intelligence. XAI research emphasizes that automated systems should provide reasoning that users and domain experts can inspect, while human–AI alignment research highlights the importance of aligning computational decisions with expert logic and collaborative decision-making (Kuznietsov et al., 2024; Sun et al., 2026).

Multimodal and human-centered applications extend these principles beyond conventional scene recognition. Vision–language models have been applied to street-level assessment, while accessibility-oriented frameworks demonstrate that objective visual measurements alone do not fully capture human interpretation of urban environments (Wang et al., 2024; Yuan et al., 2026). Related transportation systems, including shared steering and remote assistance for automated vehicles, also show that contextual awareness can improve user acceptance and human–machine collaboration (Guo et al., 2024; Schrank et al., 2024).

Similar priorities appear in GeoAI and intelligent information systems, where explainability, contextual adaptation, and integration of human expertise are increasingly emphasized (Ye et al., 2025a; Li Q. et al., 2025). Human-centered cyber-physical-social systems provide a related direction, while specialized transportation datasets with detailed road and infrastructure annotations support the development of reproducible safety-analysis methods (Alsamhi et al., 2024; Tbaileh et al., 2026).

Overall, prior work has achieved substantial progress in semantic segmentation, contextual perception, multimodal reasoning, and explainable transportation intelligence. However, most approaches remain focused on computational scene interpretation or automated decision-making. Relatively few studies combine explicit variation across Urban, Residential, and Rural environments with multidisciplinary expert judgments and transparent rules for perceived road-safety assessment. This leaves a gap between advances in visual scene understanding and interpretable frameworks designed to approximate expert safety perception across heterogeneous road contexts.

2.2 Perceived road safety assessment using street-view imagery

Street-view imagery has enabled road safety to be studied from visual, contextual, and perceptual perspectives, complementing traditional assessments based on crash records. Studies combining street-view images with semantic segmentation and explainable machine learning have shown that sidewalks, roads, buildings, vegetation, and other environmental features influence perceived safety (Hamim and Ukkusuri, 2024). At the same time, comparisons with accident data demonstrate that environments perceived as visually safe do not necessarily exhibit lower crash risk (Yu et al., 2024).

Machine-learning analyses further indicate that road configuration, environmental complexity, and roadside characteristics are associated with safety outcomes (Cai et al., 2022). Semantic urban analysis and built-environment research similarly show that perceived risk depends on the broader scene context and subjective interpretation rather than isolated objects (Liu et al., 2025). Eye-tracking research adds a cognitive dimension by demonstrating that visual attention influences how road environments are interpreted and evaluated (Kang et al., 2026).

Street-view imagery has also supported assessments focused on vulnerable road users. Deep learning and semantic segmentation have been applied to cyclist-safety evaluation, while explainable models show that perceived cycling safety depends on infrastructure configuration, green space, and environmental complexity (Ye et al., 2024; Rita et al., 2023). Related street-quality and explainable deep-learning studies confirm that multidimensional contextual characteristics are necessary to interpret human perceptions of urban environments (Wang et al., 2024; Hou and Zhao, 2024).

Recent reviews identify explainable AI, multimodal transportation intelligence, and contextual environmental reasoning as important directions for future safety assessment (Kuznietsov et al., 2024; Ye et al., 2025b). This direction is consistent with systems-based road-safety research, which treats safety as an interaction among infrastructure, environmental conditions, and human behavior, and with human-centered models emphasizing perception and behavioral interpretation (Hamim et al., 2022; Siebke et al., 2023).

The reliability of expertise-based evaluation is therefore particularly important. Expert road-safety inspections have shown meaningful associations with observed accident occurrence, while structured street-view audits have achieved high inter-rater reliability (Borsos et al., 2026; Saito et al., 2022). Multi-rater evaluation research further emphasizes the need for explicit agreement analysis when constructing expert-derived reference standards (Konstantinidis et al., 2022).

Although this literature has advanced semantic scene understanding, explainable safety assessment, and human-perception analysis, many approaches continue to prioritize segmentation performance, crash prediction, or a single evaluator perspective. Relatively little work integrates explicit Urban, Residential, and Rural context variation with multidisciplinary expert judgments and transparent semantic reasoning within one reproducible framework.

Recent work has explored MLLMs and VLMs for persona-based urban safety perception, reliability-aware comparison with human judgments, interpretable concept-driven urban perception, and safety-critical traffic control (Beneduce et al., 2025; Mushkani, 2026; Yu et al., 2025; Wang et al., 2025). More broadly, these models show promise for simulating human perception and supporting expert-aligned decision-making, while raising challenges related to reasoning consistency and explainability (Wu et al., 2026; Yin et al., 2026; Lu et al., 2026). Since perceived safety is inherently multidimensional and context-dependent (Qin et al., 2025), CARSA addresses a narrower task: expert-aligned, three-level assessment of perceived road-infrastructure safety using predefined semantic and contextual rules. It is therefore positioned not as a replacement for VLMs, but as a complementary, interpretable, and auditable approach.

3 Methodology

3.1 Image selection and semantic filtering

This study utilized the Mapillary Vistas dataset as the primary source of street-level imagery because of its geographic diversity, dense pixel-level semantic annotations, and broad representation of transportation environments (Hussain et al., 2022; Randhawa et al., 2025; Yang et al., 2022). The complete image-selection workflow is illustrated in Figure 3. Rather than randomly selecting images, a semantic-richness-guided sampling strategy was developed to construct a balanced benchmark containing scenes with different levels of transportation-related semantic complexity.

FIGURE 3

3.1.1 Road-safety semantic abstraction

The original Mapillary Vistas taxonomy contains 66 semantic categories. To reduce visual redundancy and retain information relevant to perceived road safety, these categories were consolidated into fifteen road-safety-oriented classes:

other, road, sidewalk, lane marking, curb, barrier, traffic sign, traffic light, pole, pedestrian, cyclist/rider, vehicle, vegetation, building, and crosswalk.

Each original semantic class was deterministically assigned to exactly one of these categories, while less relevant classes were grouped into the other category. The selected categories reflect semantic elements that contribute directly or indirectly to perceived transportation safety (Hamim and Ukkusuri, 2024; Wang et al., 2024; Qin et al., 2025).

After semantic abstraction, 20,000 street-level images from the official training and validation partitions of Mapillary Vistas were analyzed. The test partition was intentionally excluded to preserve compatibility with the original benchmark design.

3.1.2 Semantic richness estimation

A semantic richness score was computed for each image to quantify the diversity of transportation-related information contained in the scene. The score was intended to identify visually informative images for expert evaluation rather than estimate road safety directly.

The semantic richness score of image was defined as

where is the set of unique road-safety semantic categories present in image , and denotes the literature-informed heuristic importance weight assigned to category during benchmark construction.

Each category contributes only once to the final score irrespective of its spatial extent (Hamim and Ukkusuri, 2024; Wang et al., 2024). Higher weights were assigned to semantic categories that frequently influence human road-safety interpretation, whereas contextual background elements such as vegetation and buildings received lower weights (Hamim and Ukkusuri, 2024; Ye et al., 2024; Liu et al., 2025). The adopted weighting scheme was defined heuristically by the authors and informed by prior literature on perceived road safety and transportation environments. These weights were used only during semantic-richness-based image sampling and were unrelated to the later expert safety annotations.

It is important to distinguish the semantic-richness weighting used for benchmark construction from the context-specific weights used by CARSA. The semantic-richness score was employed exclusively to stratify the original Mapillary Vistas images according to semantic diversity and to support balanced random sampling. It was not designed as a perceived-safety score, was not derived from expert annotations, and was explicitly removed before CARSA feature extraction and safety-score computation. Consequently, the image-selection criterion did not contribute to the framework predictions.

In addition to the weighted semantic richness score, the number of critical road-safety semantic categories present in each image was recorded as an auxiliary descriptor of scene complexity. These categories included sidewalks, lane markings, barriers, traffic signs, traffic lights, pedestrians, cyclists/riders, vehicles, and pedestrian crossings.

3.1.3 Semantic-richness stratification and balanced sampling

After computing the semantic richness score using (Equation 1), the score distribution was partitioned into three semantic-richness strata to ensure balanced representation of scenes with different levels of semantic complexity.

Let denote the semantic-richness level assigned to image :

The thresholds were determined empirically from the semantic richness distribution and were introduced exclusively for balanced dataset construction.

To avoid over-representation of any semantic-richness level, images were first grouped according to their semantic-richness category and sampled independently. Specifically, 140 images were randomly selected from each of the High-, Medium-, and Low-semantic-richness groups using a fixed random seed (42). The use of a fixed seed ensures complete experimental reproducibility and enables independent reconstruction of the benchmark dataset. This procedure resulted in a balanced subset of 420 images. The sampled images were subsequently shuffled using the same random seed before any manual inspection or expert evaluation.

3.1.4 Human-relevance filtering

Although semantic-richness-guided sampling ensured balanced semantic complexity, semantic richness alone does not guarantee that a scene is suitable for human-centered road-safety evaluation. Therefore, the balanced subset of 420 images was subjected to an additional manual human-relevance filtering stage.

Images were excluded based on the inability to perform reliable multidisciplinary safety assessment using the visual evidence contained within the image itself rather than because they represented unsafe environments.

The filtering stage was not intended to remove scenes merely because their perceived-safety interpretation was difficult, ambiguous, or likely to produce disagreement. All scenes for which the experts could form a safety judgment from the available visual evidence were retained, including borderline and context-sensitive cases. Exclusion was limited to scenes for which the intended human-centered assessment could not be performed reliably because the relevant road interaction, infrastructure, or environmental context was not sufficiently observable.

Applying these criteria reduced the dataset from 420 images to a final benchmark containing 315 images, while 105 images were excluded as being outside the intended scope of the framework. Representative examples are shown in Figure 4.

FIGURE 4

The complete lists of retained and excluded images have been publicly released together with the source code and experimental resources through the Data and Code Availability section, enabling independent verification of the dataset construction procedure and supporting full experimental reproducibility.

The proposed dataset construction framework combines semantic abstraction, weighted semantic-richness estimation, stratified sampling, and human-centered filtering to generate a balanced benchmark specifically designed for multidisciplinary perceived road-safety assessment.

3.2 Context-aware road categorization

Road-safety interpretation is inherently context dependent. Although semantic scene-understanding techniques can identify roads, sidewalks, traffic signs, vehicles, pedestrians, and other transportation-related objects, the safety implications of these elements may vary according to the surrounding environment. Research on perceived safety and street-level environmental assessment shows that safety judgments depend on broader contextual characteristics rather than isolated visual objects (Yu et al., 2024; Hamim and Ukkusuri, 2024; Wang et al., 2024; Liu et al., 2025). Studies focused on vulnerable road users similarly demonstrate that infrastructure elements must be interpreted within their functional road context (Ye et al., 2024; Rita et al., 2023).

To explicitly incorporate contextual information, every image retained after semantic filtering was manually assigned to one of three road-environment categories: Urban, Residential, or Rural.

The assigned context remained fixed throughout all subsequent stages and served as an explicit input variable for the Context-Aware Road Safety Assessment (CARSA) framework. Categorization was performed through manual visual inspection and completed before expert safety evaluation to prevent any influence of the final consensus labels.

Urban environments were generally characterized by dense transportation infrastructure, mixed or commercial land use, higher traffic complexity, and a greater concentration of vulnerable road users and traffic-control elements. Residential environments typically consisted of local access roads, lower traffic density, residential housing, and reduced transportation complexity. Rural environments generally exhibited lower infrastructure density, larger open spaces, and fewer traffic-control elements.

The introduction of contextual categorization was motivated by the observation that identical semantic features may carry different safety implications across transportation environments. For example, the absence of sidewalks or pedestrian crossings may indicate inadequate infrastructure in dense urban areas but may be functionally acceptable in certain rural settings.

More detailed functional classifications, such as arterial, collector, local, or expressway roads, require additional information regarding road hierarchy, traffic volume, roadway function, speed regulations, and network connectivity that cannot be reliably inferred from isolated street-level imagery. Consequently, grouping scenes into Urban, Residential, and Rural environments provides an appropriate balance between contextual specificity and robust visual interpretation.

The final dataset consisted of 231 Urban images, 42 Residential images, and 42 Rural images. Representative examples are presented in Figure 5.

FIGURE 5

3.3 Context-Aware Road Safety Assessment framework

3.3.1 Label-independent development and parameter freezing

CARSA was developed as a deterministic, rule-based framework rather than as a supervised model fitted to perceived-safety labels. The development sequence was explicitly separated from the subsequent human evaluation. First, prior research on perceived road safety, road infrastructure, vulnerable-road-user protection, and context-dependent transportation assessment was reviewed to identify the relevant semantic classes and their functional grouping into pedestrian support, traffic control, and physical protection factors. Second, the presence–coverage fusion coefficients, context-specific factor weights, scoring equations, and categorical decision thresholds were predefined. Third, the resulting implementation was frozen and used to generate the CARSA scores and categorical predictions for all benchmark images.

These stages were completed before the expert annotation process. At the time the CARSA parameters and predictions were finalized, no individual expert ratings, consensus labels, confusion matrices, agreement statistics, or framework-performance results were available. The expert annotations were subsequently used only as an independent reference standard for evaluation. No rule, coefficient, factor weight, normalization procedure, or decision threshold was selected, optimized, recalibrated, or revised using the expert ratings.

The framework-generation notebook and the human-evaluation notebook were maintained as separate computational stages. The former was used to generate the CARSA predictions without loading any expert-rating file, whereas the latter imported the frozen predictions and expert annotations only for statistical comparison. This separation provides an auditable workflow and prevents target-label information from entering the framework-prediction process.

The use of the same 315 images by CARSA and the expert panel was a necessary component of the paired evaluation design rather than a source of information leakage. Quantifying image-level agreement between the framework and human judgment requires both CARSA and the experts to assess the identical road scenes; otherwise, the framework predictions could not be compared with a corresponding human Ground Truth.

The methodological independence of the evaluation was therefore established through complete temporal and informational separation between the two assessment processes, rather than through the use of different images. The images served as common evaluation units, while the expert-derived perceived-safety labels served exclusively as the independent Ground Truth. These labels were created only after the CARSA parameters had been frozen and the framework predictions had been generated and stored. They were never used to select, adjust, optimize, or recalibrate the CARSA rules, factor weights, normalization procedure, or decision thresholds, as shown in Table 1.

3.3.2 Semantic feature extraction and safety-score computation

After assigning each image to a road-environment context, CARSA was applied under the frozen framework specification described in Section 3.3.1 to translate semantic scene information into an interpretable perceived-safety score. Expert ratings were neither available nor used during semantic-feature construction, parameter definition, score computation, or categorical prediction.

The framework operates on the final semantically filtered dataset and uses the road-context label assigned in Section 3.2. The weighted semantic-richness score used during benchmark construction was not used as an input to CARSA. It was employed exclusively during the earlier image-sampling stage and was removed from the working metadata before semantic-feature extraction and safety-score computation.

The semantic masks used in this study are the official Mapillary Vistas annotations and are treated as fixed inputs to CARSA. The framework does not perform semantic segmentation or object detection; the present evaluation therefore isolates the downstream contextual reasoning stage from the accuracy and design choices of any upstream perception model.

For each image , the semantic segmentation mask was processed to extract binary semantic presence and normalized semantic coverage. Let denote the reduced set of road-safety semantic classes:

For each semantic class , a binary presence indicator was defined as

The raw pixel coverage ratio was computed as

where denotes the number of pixels belonging to semantic class in image , and denotes the total number of pixels.

Each coverage value was normalized independently using the maximum observed coverage of that class:This normalization uses only unlabeled pixel-level semantic-mask statistics and does not use individual expert ratings, consensus labels, or perceived-safety categories. It was computed before the expert evaluation and therefore does not represent label-based calibration. The normalization was intentionally restricted to the human-relevant benchmark domain because the complete Mapillary pool contains controlled-access highways, isolated corridors, highly occluded scenes, and other images outside the intended scope of the framework. The resulting coverage normalization should nevertheless be interpreted as benchmark-relative rather than as a universal cross-dataset normalization.

To combine semantic availability and spatial prominence, a presence–coverage indicator strength was computed as

The larger weight assigned to binary presence reflects the assumption that human perceived road safety depends primarily on whether safety-relevant infrastructure exists, while normalized coverage provides complementary information about its visual prominence.

Although pedestrians, cyclists, and vehicles were extracted as descriptive semantic classes, they were not used directly in the final safety-score computation because the framework aims to assess perceived infrastructure safety rather than transient traffic conditions.

Three infrastructure-based safety factors were computed:

Here, represents pedestrian-support infrastructure, represents traffic-control and guidance elements, and represents physical protection and separation.

To incorporate road-context dependency, the final continuous safety score was computed as

where denotes the context-aware perceived safety score and denotes the road-environment context of image . The context-specific weights are summarized in Table 2.

Finally, the continuous CARSA score was converted into categorical perceived safety levels according to

The context-specific thresholds are reported in Table 3.

3.3.3 Parameter rationale and threshold interpretation

The CARSA parameters were predefined heuristic operational values rather than coefficients estimated from expert labels. The road-safety literature informed the selection and contextual importance of the three infrastructure factors, whereas the numerical values were chosen to preserve a transparent and interpretable scoring structure.

The context-specific factor weights reflect differences in the expected function of infrastructure across road environments. Urban scenes assign the highest importance to pedestrian-support and traffic-control infrastructure because of denser interaction among pedestrians, vehicles, and control devices. Residential scenes emphasize pedestrian support and physical separation, while rural scenes assign greater importance to physical protection and traffic guidance because dedicated pedestrian infrastructure is less consistently expected.

The categorical thresholds were adopted as predefined heuristic operating points before expert annotation. Their numerical positions can be interpreted using presence-based reference configurations consistent with the CARSA scoring equation. Since semantic presence contributes 0.80 to each indicator strength, a three-element factor contributes approximately 0.267, 0.533, or 0.800 when one, two, or all three constituent elements are present, respectively, before the additional coverage contribution. A two-element protection factor contributes approximately 0.400 or 0.800 when one or both elements are present. These configurations demonstrate that the adopted thresholds are structurally consistent with interpretable combinations of safety-related infrastructure; they do not imply that the thresholds were statistically estimated from expert labels, as shown in Table 4.

These configurations are illustrative rather than exclusive, because multiple combinations of semantic indicators may produce the same continuous CARSA score. Importantly, the original thresholds were fixed before expert annotation and were not subsequently modified after the human–framework agreement results became available.

3.3.4 Framework outputs

The final framework output for each image consisted of the road context, the continuous safety score , the three interpretable safety-factor scores, and the categorical safety label . These outputs were generated and stored before the expert annotations were introduced into the evaluation workflow.

Equations 3–12, together with the context-specific weights in Table 2 and the decision thresholds in Table 3, constitute the complete CARSA rule specification. No additional learned parameters, class-specific overrides, hidden decision rules, post-processing steps, or exceptions were used to generate the reported CARSA predictions.

Algorithm 2 summarizes the complete deterministic inference procedure used to generate the reported CARSA factor scores, continuous safety scores, and categorical predictions.

3.3.5 Predefined parameter-sensitivity analysis

To examine whether the reported CARSA performance depended excessively on a single parameter specification, a predefined one-at-a-time sensitivity analysis was conducted around the frozen CARSA v1 configuration. The analysis retained the original framework as the reference configuration and evaluated 64 alternative configurations, resulting in 65 configurations in total. All alternative predictions were generated and exported before loading the expert annotations; therefore, the sensitivity experiment was not used to tune, recalibrate, or replace the original CARSA parameters.

Four parameter groups were examined. First, the presence–coverage fusion was evaluated using presence-to-coverage weights of and , in addition to the frozen baseline. Second, each context-specific factor weight was perturbed individually by and , while the remaining weights were proportionally renormalized so that the three factor weights continued to sum to one. This produced 36 factor-weight configurations across the Urban, Residential, and Rural contexts. Third, each Low–Medium and Medium–High decision boundary was shifted independently by and , producing 24 threshold configurations while preserving the ordering of the two thresholds.

Finally, the benchmark-wide maximum used for semantic-coverage normalization was compared with two five-fold cross-fitted alternatives. The folds were stratified by road context using a fixed random seed of 42. For each held-out fold, normalization constants were estimated exclusively from the other four folds using either the maximum or the 95th percentile of all training-fold coverage values, including zeros. Normalized values exceeding one were clipped to one.

Each configuration was evaluated against the independently constructed expert consensus using accuracy, weighted precision, weighted recall, weighted F1, Cohen’s Kappa, quadratic weighted Kappa, Pearson correlation, Spearman correlation, context-specific accuracy, and categorical stability relative to the frozen baseline. All configurations were retained in the analysis, regardless of whether their performance increased or decreased. Consequently, the experiment should be interpreted as a parameter-sensitivity assessment on the existing benchmark rather than as held-out parameter optimization or external validation, as shown in Table 5.

3.4 Expert evaluation protocol

Road safety assessment extends beyond visual object recognition and involves the interpretation of infrastructure design, road-user interaction, environmental context, and professional judgment (Hamim et al., 2022; Siebke et al., 2023). To establish a human-centered reference standard for perceived road safety, the final dataset of 315 street-level images was independently evaluated by a multidisciplinary expert panel.

The panel consisted of three specialists with complementary expertise in transportation engineering, practical road-use assessment, and computer vision-based scene interpretation. This multidisciplinary composition was intended to reduce the limitations associated with a single professional perspective while capturing multiple dimensions of road-safety perception.

The transportation engineering expert contributed knowledge related to roadway design, infrastructure quality, and traffic organization. The driving instructor contributed practical expertise regarding driver behavior, visibility conditions, conflict exposure, and real-world traffic interactions. The computer vision researcher contributed expertise in semantic scene understanding and transportation-image interpretation. Integrating these complementary viewpoints enabled a broader and more human-centered interpretation of road environments.

The use of multidisciplinary expert annotation was further motivated by the observation that perceived road safety often depends on professional background and domain knowledge. Previous studies have shown that expert-annotated datasets provide more reliable benchmark references than automatically inferred labels when the target concept requires substantial domain knowledge and cannot be directly observed from raw data alone (Joyce et al., 2023; Willemink et al., 2023).

Each image was evaluated independently by all three experts using an ordinal three-level perceived road-safety scale consisting of Low Safety, Medium Safety, and High Safety. Experts were not provided with the evaluations of the other panel members during the labeling procedure.

The experts were also blinded to the CARSA outputs during the annotation process. In particular, they were not shown the continuous CARSA safety scores, factor-level scores, or predicted Low-, Medium-, and High-safety labels. The framework predictions had already been generated and stored, but they were not introduced into the human-evaluation workflow until all individual expert ratings had been completed.

The assessment was based exclusively on the visual information available within each image and its surrounding environmental context rather than on external information such as crash statistics, traffic volume measurements, roadway inventory databases, or automated model predictions. Experts were instructed to evaluate the overall safety characteristics represented in each scene by considering infrastructure quality, road organization, pedestrian and cyclist protection, visibility conditions, traffic-control devices, vehicle interaction, conflict exposure, and the suitability of the transportation environment within its corresponding context. No CARSA-derived score or categorical prediction was displayed during annotation, thereby preserving the independence of the human reference labels from the framework outputs.

High Safety labels were assigned to scenes exhibiting well-organized transportation infrastructure, clear traffic organization, adequate protection for vulnerable road users, and relatively limited exposure to potential conflicts. Medium Safety labels represented environments with partially adequate infrastructure or moderate safety concerns. Low Safety labels were assigned to scenes characterized by poor infrastructure support, inadequate road organization, elevated conflict exposure, or limited protection for vulnerable road users.

Unlike conventional computer vision approaches that focus on isolated objects, the proposed evaluation protocol emphasized holistic scene understanding, where experts interpreted the complete transportation environment rather than evaluating individual semantic elements independently (Yu et al., 2024; Hamim and Ukkusuri, 2024; Saito et al., 2022).

To facilitate subsequent statistical analysis, the ordinal safety categories were converted into numerical values according to

where denotes the rating assigned by expert to image , with one corresponding to Low Safety, 2 to Medium Safety, and three to High Safety. Since each of the 315 images was evaluated by all three experts, the final annotation dataset contained 945 independent evaluations.

The individual expert ratings were organized into an expert-rating matrix. For each image, a continuous human consensus score was computed as

where denotes the mean human consensus score for image (i). The degree of disagreement among experts was quantified using the population standard deviation:

where represents the expert disagreement score associated with image (i).

Finally, the continuous consensus score was converted back into a categorical human reference label according to

Applying this procedure resulted in 200 High Safety images (63.49%), 95 Medium Safety images (30.16%), and 20 Low Safety images (6.35%). These consensus labels served as the reference standard for all subsequent framework evaluation and statistical analyses.

3.5 Comparative baselines and statistical analysis

To contextualize CARSA’s performance, four supporting baselines were evaluated on the same 315-image benchmark. First, a non-contextual equal-weight semantic baseline was constructed as an ablation of CARSA. It retained the same pedestrian-support, traffic-control, and physical-protection factors, but removed the context-specific weighting and decision rules. The three factors were averaged equally, and fixed global thresholds of 0.40 and 0.70 were used to obtain the Low, Medium, and High safety categories.

A context-aware zero-shot CLIP baseline was also implemented using the pretrained CLIP ViT-B/32 model (openai/clip-vit-base-patch32). Fixed prompt ensembles were defined for the three safety categories within the Urban, Residential, and Rural contexts. The predicted category corresponded to the context-specific text prototype with the highest cosine similarity to the image embedding. The semantic and CLIP baselines were evaluated without training, fine-tuning, parameter fitting, or access to the expert-consensus labels during prediction generation.

The comparison additionally included Qwen2.5-VL-7B-Instruct under zero-shot and few-shot prompting. The zero-shot configuration relied on task instructions only, whereas the few-shot configuration used nine fixed in-context demonstrations covering the three road contexts and safety levels. No parameter updating or fine-tuning was performed.

All methods were evaluated against the same independent expert consensus using accuracy, balanced accuracy, weighted F1, macro F1, Cohen’s Kappa, Quadratic Weighted Kappa, and ordinal mean absolute error. Differences in paired classification accuracy between CARSA and each baseline were assessed using exact McNemar tests. Paired bootstrap confidence intervals were estimated using 10,000 resamples, and Holm correction was applied across the four statistical comparisons.

3.6 Human evaluation reliability analysis

Before evaluating the proposed framework, the reliability and consistency of the human-generated labels were assessed. Since road-safety perception involves subjective judgment and multidisciplinary interpretation, establishing a reliable expert consensus is necessary to ensure the validity of subsequent human–AI agreement analyses.

Several complementary reliability measures were employed to evaluate inter-rater reliability, internal consistency, categorical agreement beyond chance, and ordinal consistency.

The Intraclass Correlation Coefficient (ICC) was used to evaluate the overall consistency of ratings among the three experts. The obtained ICC(A,k) value was 0.953, indicating excellent reliability.

Internal consistency was examined using Cronbach’s Alpha. The resulting value of 0.954 also indicates excellent reliability and confirms highly consistent rating behavior.

To evaluate categorical agreement beyond chance, Fleiss’ Kappa was computed across the three safety categories. The resulting value of 0.805 indicates almost perfect agreement according to the interpretation guidelines proposed by Landis and Koch.

In addition, the percentage of complete agreement was calculated to measure the proportion of images for which all three experts assigned the same safety category. Complete agreement was achieved for 85.08.

Overall, the reliability analysis demonstrates that the expert-generated perceived road-safety labels provide a stable and trustworthy reference standard for evaluating the proposed framework.

The distribution of the expert-consensus safety classes across the three road contexts is reported in Table 6, together with the class-specific precision, recall, and F1 scores achieved by CARSA.

The benchmark was dominated by High-safety scenes, particularly within the Urban context. CARSA achieved its highest class-specific performance for the High-safety category, with an F1 score of 0.9614. The Low-safety class had the smallest support and the lowest precision, although its recall remained high at 0.9000. These class-specific results complement the context-specific performance reported in Table 9.

4 Results

4.1 Overall human–framework agreement analysis

The overall agreement analysis was conducted to evaluate the extent to which the proposed CARSA framework could reproduce the expert consensus labels across the complete benchmark of street-level images.

Table 7 summarizes the resulting performance. The proposed framework achieved an overall classification accuracy of 92.38%, indicating strong agreement with the multidisciplinary expert consensus. This result suggests that the semantic and context-aware scoring mechanism successfully captures many of the visual and contextual factors that influence human road-safety perception. Given the inherently subjective nature of perceived safety assessment, the observed level of agreement demonstrates that the framework can reliably approximate expert judgment across diverse transportation environments.

Beyond overall accuracy, agreement beyond chance was evaluated using Cohen’s Kappa. The framework achieved a Cohen’s Kappa value of 0.8529, indicating almost perfect agreement with the expert consensus. This result demonstrates that the observed performance cannot be attributed solely to the class distribution of the dataset, but instead reflects a strong correspondence between framework predictions and human safety judgments.

To account for the ordinal nature of the perceived safety labels, Quadratic Weighted Kappa (QWK) was also computed. The resulting QWK value of 0.8914 further confirms the strong alignment between framework outputs and expert evaluations. The higher QWK value relative to the unweighted Kappa suggests that most disagreements occurred between neighboring safety categories rather than between opposite extremes. This interpretation is consistent with the failure severity analysis presented in Section 4.4, where only a single image exhibited the maximum possible disagreement across the three safety levels. Collectively, these findings indicate that even when classification errors occurred, the framework generally remained close to the corresponding human judgment.

The framework also demonstrated strong score-level alignment with human evaluations. The Pearson correlation coefficient reached 0.8246, indicating a strong linear relationship between the continuous framework scores and the expert-derived safety ratings. Similarly, the Spearman rank correlation coefficient reached 0.7968, demonstrating that the framework largely preserved the relative ordering of road scenes according to perceived safety.

The results suggest that the framework captures the underlying structure of human safety perception and provides meaningful continuous safety estimates.

Figure 6 provides a detailed visualization of the agreement structure between framework predictions and expert consensus labels. A clear concentration of observations along the main diagonal can be observed, confirming strong agreement across all three perceived safety categories. The framework correctly classified 187 of 200 High-safety scenes, 86 of 95 Medium-safety scenes, and 18 of 20 Low-safety scenes.

FIGURE 6

Most errors were confined to adjacent safety categories rather than opposite extremes, indicating that disagreements were primarily associated with borderline or context-sensitive cases. The near absence of severe errors further suggests that the framework largely preserved the ordinal structure of human safety perception even when exact agreement was not achieved.

4.1.1 Parameter-sensitivity results

Table 8 summarizes the results of the predefined parameter-sensitivity analysis. The frozen CARSA v1 configuration achieved an accuracy of 92.38%, a weighted F1 score of 92.61%, Cohen’s Kappa of 0.853, and a quadratic weighted Kappa of 0.891. The sensitivity results indicate that the framework was generally stable under moderate changes to the context-specific factor weights and semantic-coverage normalization. Across the 36 factor-weight perturbations, accuracy ranged from 87.94% to 93.65%, while categorical stability remained between 94.92% and 100%. The two cross-fitted normalization alternatives produced accuracies between 92.70% and 93.65% and preserved between 96.83% and 99.05% of the baseline labels.

The presence–coverage experiment revealed an asymmetric response. Increasing the presence contribution from 0.80 to 0.90 produced an accuracy of 94.60% with 95.87% label stability, whereas increasing the coverage contribution from 0.20 to 0.30 reduced accuracy to 68.25% and label stability to 74.29%. This indicates that the framework is more sensitive to increasing the influence of spatial coverage than to further emphasizing semantic presence. The higher performance of the alternative is reported only as a sensitivity result and was not used to replace the predefined baseline.

Decision-threshold perturbations produced the widest performance range. Most individual threshold changes affected only a small number of images; however, raising the Urban Medium–High boundary from 0.70 to 0.73 or 0.75 changed 52 and 55 Urban predictions, respectively, from High to Medium. These two configurations reduced overall accuracy to 76.51% and 75.56%, demonstrating localized sensitivity around the upper Urban decision boundary. In contrast, lowering this boundary to 0.67 or 0.65 changed only five predictions and produced an accuracy of 93.33%. Because threshold perturbations alter only the categorical assignment and not the continuous CARSA score, Pearson and Spearman correlations remained unchanged within this parameter group.

Importantly, none of the 64 alternative configurations produced a two-level categorical transition relative to the frozen baseline; all changed predictions moved only between adjacent safety levels. Overall, the results support the local stability of CARSA under moderate factor-weight and normalization perturbations, while identifying the presence–coverage balance and the Urban Medium–High boundary as the most influential parameter components.

4.2 Road-type-specific performance analysis

To further investigate the effectiveness of the proposed framework across different transportation environments, a road-type-specific analysis was conducted. Because the framework explicitly incorporates contextual reasoning through urban, residential, and rural categorization, evaluating performance within each road context provides important insights into its context-specific robustness within the current benchmark.

Table 9 summarizes the framework performance across the three road environments. Performance varied considerably across contexts, with the strongest agreement observed in urban roads, followed by rural roads, while residential scenes proved comparatively more challenging.

The highest performance was observed in urban environments, where the framework achieved an accuracy of 96.54%. Urban scenes typically contain rich semantic information, including traffic-control devices, road markings, pedestrian infrastructure, and diverse transportation-related objects. The abundance of contextual cues likely contributed to the strong agreement between framework predictions and expert evaluations.

Rural environments also demonstrated strong performance, with an accuracy of 90.48%. Although rural scenes typically contain fewer explicit safety-related elements than urban environments, their visual structure and contextual characteristics appear to be more consistently represented within the proposed scoring framework. As a result, a high level of agreement was maintained between framework predictions and expert evaluations.

The lowest performance was observed in residential environments, where the framework achieved an accuracy of 71.43%. Compared with urban and rural scenes, residential environments often contain more ambiguous combinations of safety-related cues, making perceived safety assessment inherently more challenging. In many cases, the distinction between Medium- and High-safety conditions may depend on subtle contextual characteristics that are difficult to represent using visible infrastructure features alone. Consequently, disagreements between framework predictions and expert assessments occurred more frequently within this category.

Overall, the results highlight the importance of contextual reasoning in perceived road-safety assessment, indicating that human safety perception depends not only on infrastructure elements but also on the environmental context in which they appear.

Figure 7 further illustrates the performance differences summarized in Table 9, with residential scenes remaining the most challenging context for human-aligned safety assessment.

FIGURE 7

These performance differences suggest that perceived road safety is strongly influenced by environmental context. In particular, residential scenes often contain more ambiguous safety cues, resulting in more frequent disagreements between framework predictions and expert judgments.

Despite these variations, the framework maintained strong agreement in the Urban and Rural subsets, while lower agreement was observed in the Residential subset. These results support the effectiveness of context-aware reasoning across the three road environments represented within the current benchmark, while cross-dataset and cross-geographic generalization remain to be examined in future work.

4.3 Human consensus and framework error analysis

To better understand the conditions under which the proposed framework succeeded or failed, the relationship between human expert agreement and framework prediction errors was examined. This analysis is particularly important because perceived road-safety assessment is inherently subjective, and disagreements among experts may indicate images that are intrinsically more difficult to interpret.

The evaluation dataset contained 315 images, of which 291 were correctly classified and 24 were misclassified by the framework, corresponding to an overall error rate of 7.62%. Human disagreement was quantified using the standard deviation of expert ratings for each image. For correctly classified images, the average disagreement score was 0.0632, indicating a high level of consensus among experts. In contrast, images misclassified by the framework exhibited a substantially higher average disagreement score of 0.1571.

This finding suggests that framework errors were not distributed randomly across the dataset. Instead, misclassifications occurred more frequently in scenes that also exhibited greater uncertainty among human evaluators and were associated with inherently ambiguous or context-sensitive road environments.

Consistent with this observation, full human agreement was present in 86.60% of correctly classified images, compared with only 66.67% of misclassified images. These results indicate that the framework performed most reliably in road scenes with high expert consensus.

To further examine this relationship, a Mann–Whitney U test was conducted to compare human disagreement levels between correctly classified and misclassified images. The test produced a U statistic of 2796.0 and a p-value of 0.0086, indicating a statistically significant difference between the two groups. These results provide quantitative evidence that framework errors are associated with higher levels of human uncertainty and suggest that many apparent failures occurred in scenes that were inherently difficult for human evaluators to assess consistently, reinforcing the subjective and context-dependent nature of perceived road-safety assessment.

Overall, the results suggest that framework errors were concentrated in road scenes with lower expert consensus, indicating that the proposed framework remains broadly aligned with expert judgment even in challenging evaluation conditions. Table 10 summarizes the differences in expert agreement characteristics between correctly classified and misclassified images.

4.4 Failure severity analysis

While the overall framework performance was strong, examining the severity of prediction errors provides additional insight into the practical reliability of the proposed approach. Because perceived road-safety assessment is inherently ordinal, disagreements between adjacent safety categories are less consequential than disagreements between opposite extremes.

To evaluate the practical significance of framework errors, prediction disagreements were analyzed according to their distance from the corresponding human consensus labels, providing a more nuanced assessment of framework behavior than overall accuracy alone.

Table 11 summarizes the distribution of prediction error distances between framework outputs and expert consensus labels. Exact agreement was achieved for 291 images (92.38%), demonstrating that the framework reproduced expert judgments for the vast majority of evaluated road scenes. Moderate disagreements, corresponding to a one-level difference between framework and human evaluations, occurred in only 23 images (7.30%). Importantly, only a single image (0.32%) exhibited the maximum possible disagreement, where the framework predicted the opposite safety category from the expert consensus.

These findings indicate that most misclassifications were confined to adjacent safety categories, further supporting the high Quadratic Weighted Kappa and demonstrating that the framework largely preserved the ordinal structure of human safety perception.

Figure 8 illustrates the distribution of disagreement severity between framework predictions and expert consensus labels. Exact agreement predominated across the dataset, while most disagreements were limited to adjacent safety categories and severe errors were exceptionally rare.

FIGURE 8

4.5 Framework bias analysis

Beyond overall agreement, it is important to examine whether the proposed framework exhibits systematic tendencies toward overestimating or underestimating perceived road safety, as such tendencies can influence the practical interpretation of framework outputs.

To investigate this aspect, framework predictions were compared with expert consensus labels and categorized into three groups: matched predictions, conservative predictions, and optimistic predictions. Conservative predictions correspond to cases in which the framework assigned a lower safety category than the expert consensus, whereas optimistic predictions correspond to cases in which the framework assigned a higher safety category.

The results are presented in Table 12. Exact agreement between the framework and expert consensus labels was observed for 291 images (92.38%). Among the remaining cases, 20 images (6.35%) were classified more conservatively than the expert consensus, whereas only four images (1.27%) were classified more optimistically.

These findings indicate a modest conservative tendency in the framework predictions, with disagreements occurring more frequently as underestimations of perceived safety than as overestimations. Nevertheless, the magnitude of this bias remained small relative to the overall agreement rate.

Figure 9 illustrates the distribution of framework prediction bias relative to expert consensus labels. Exact agreement predominated across the dataset, while the relatively small number of disagreements showed a modest tendency toward conservative predictions.

FIGURE 9

4.6 Road-type bias analysis

The previous analysis identified a modest conservative tendency in the overall framework predictions. To investigate whether this behavior varied across transportation environments, directional prediction bias was examined separately for urban, residential, and rural road contexts.

Table 13 summarizes the distribution of prediction outcomes across the three road categories. Consistent with the overall performance analysis, urban environments exhibited the strongest agreement with expert consensus labels and the lowest proportion of prediction errors. In contrast, residential environments displayed a noticeably higher frequency of conservative disagreements, indicating that the framework was more likely to underestimate perceived safety in these contexts. Rural roads demonstrated comparatively stable behavior, with only a limited number of conservative and optimistic disagreements.

These findings suggest that prediction bias is influenced by environmental context, with residential roads remaining the most challenging setting for human-aligned safety assessment.

The concentration of conservative predictions in residential environments suggests that certain contextual factors influencing human safety perception are not fully captured by visible infrastructure characteristics alone. This observation is consistent with the lower agreement rates observed for residential roads in the road-type-specific performance analysis.

Overall, the results indicate that environmental context influences both framework performance and prediction bias, supporting the need for context-aware approaches to perceived road-safety assessment.

4.7 Correlation between human ratings and framework scores

Beyond categorical agreement, the relationship between continuous framework scores and human safety perception was investigated using correlation analysis. For this purpose, the mean expert rating for each image was calculated by averaging the evaluations provided by the three experts (Low = 0, Medium = 1, High = 2). The resulting mean human rating was then compared with the corresponding continuous framework safety score.

The Pearson correlation coefficient reached 0.8246, indicating a strong positive linear relationship between framework outputs and expert evaluations. Similarly, the Spearman rank correlation coefficient reached 0.7968, demonstrating strong agreement in the relative ordering of road scenes according to perceived safety. These results indicate that the framework captures both the continuous and ordinal structure of human safety perception.

Figure 10 illustrates the strong positive association between mean human ratings and framework safety scores, consistent with the high Pearson and Spearman correlation coefficients.

FIGURE 10

4.8 Failure case analysis

Although the proposed framework achieved strong overall agreement with expert evaluations, examining the largest disagreement cases provides valuable insight into the limitations of the current scoring mechanism and highlights directions for future improvement.

Across the entire dataset, the average disagreement magnitude between framework scores and mean human ratings was 0.195, while the maximum observed disagreement reached 0.7455. Residential environments were disproportionately represented among the largest disagreement cases, further supporting the earlier finding that residential roads constitute the most challenging context for human-aligned safety assessment.

The most severe disagreement occurred in a residential-road scene that received unanimous expert agreement as a High-safety environment, whereas the framework classified it as Low safety. Notably, this case represented the only maximum-distance disagreement observed in the entire dataset.

An additional observation emerged from the failure analysis. Several of the largest disagreement cases occurred in images for which expert disagreement was zero, indicating complete consensus among evaluators. This finding suggests that not all framework errors can be attributed to uncertainty in human perception. Instead, some failures appear to reflect limitations of the current framework representation, particularly when experts rely on contextual cues that are not fully captured by the semantic and infrastructure-based factors incorporated into the scoring mechanism.

Overall, the failure analysis suggests that the remaining framework limitations are primarily associated with contextual information that extends beyond the current semantic representation, providing useful guidance for future framework refinement.

4.9 Comparison with semantic, CLIP, and VLM baselines

Table 14 compares CARSA with the non-contextual semantic ablation, context-aware CLIP zero-shot, and Qwen2.5-VL zero-shot and few-shot configurations. The comparison was conducted against the same independent expert consensus for all methods.

CARSA achieved the strongest agreement with the expert consensus, obtaining 92.38% accuracy, 92.61% weighted F1, 87.54% macro F1, and a QWK of 0.8914. The non-contextual semantic baseline achieved 75.87% accuracy and 77.51% weighted F1, indicating that the shared semantic representation was informative even without context-specific reasoning. Nevertheless, CARSA improved accuracy over this ablation by 16.51 percentage points, with a paired bootstrap 95% confidence interval of [12.06, 21.27] percentage points.

The CLIP zero-shot, Qwen2.5-VL zero-shot, and Qwen2.5-VL few-shot configurations achieved accuracies of 19.68%, 20.32%, and 29.84%, respectively. CARSA exceeded these baselines by 72.70, 72.06, and 62.54 percentage points, with corresponding 95% confidence intervals of [67.30, 77.78], [66.35, 77.46], and [56.51, 68.57] percentage points. All four paired improvements remained statistically significant after Holm correction in the exact McNemar tests (all adjusted ).

The Qwen2.5-VL zero-shot configuration produced no High-safety predictions, whereas its few-shot configuration assigned 314 of the 315 images to the Medium category. These concentrated prediction distributions explain the low weighted and macro F1 scores. Under the fixed protocols evaluated in this study, the general-purpose CLIP and Qwen2.5-VL representations did not reliably separate the three perceived-safety levels. These findings should not be interpreted as a general limitation of vision–language models, but rather as evidence that this task remains challenging without domain-specific adaptation or alignment.

5 Discussion

This study examined whether perceived road-infrastructure safety can be approximated through a frozen and interpretable framework based on semantic scene information and explicit road-context reasoning. The results indicate that CARSA captured substantial agreement with the multidisciplinary expert consensus on the current benchmark while preserving a transparent connection between semantic inputs, contextual weights, and categorical decisions.

The reliability of the expert reference standard is central to this interpretation. The high inter-rater agreement observed across the three professional perspectives indicates that the consensus labels represent a stable benchmark rather than an aggregation of highly inconsistent individual judgments. This finding is consistent with previous work emphasizing the importance of structured multi-rater protocols and agreement analysis in expert-derived evaluation datasets (Saito et al., 2022; Konstantinidis et al., 2022).

A particularly important result is the difference between CARSA and the non-contextual semantic ablation. Both methods used the same frozen semantic-factor construction, but CARSA improved accuracy by 16.51 percentage points, with a paired bootstrap 95% confidence interval of [12.06, 21.27] percentage points and a Holm-adjusted exact McNemar -value below . This comparison provides direct evidence that the context-specific factor weights and decision thresholds added value beyond the underlying semantic representation.

The score-level agreement and error structure further support this interpretation. Most disagreements occurred between adjacent safety categories, indicating uncertainty near decision boundaries rather than frequent reversals between Low and High safety. Nevertheless, the residential context remained substantially more challenging than the Urban and Rural contexts. This pattern suggests that residential safety judgments may depend on contextual influences not fully represented by visible infrastructure alone, such as neighborhood familiarity, tranquility, perceived driver behavior, and environmental quality. Related street-view studies similarly emphasize the interaction between physical scene characteristics and subjective contextual perception (Hamim and Ukkusuri, 2024; Ye et al., 2024). Broader urban-perception research also identifies neighborhood and environmental characteristics as important components of perceived safety (Yu et al., 2024; Wang et al., 2024; Liu et al., 2025).

CARSA exhibited a modest conservative tendency, with most misclassifications assigning a lower safety category than the expert consensus. Such behavior reduces the frequency of optimistic safety assessments, but it may also underestimate scenes that experts regard as adequately safe, particularly in residential settings. The directional bias should therefore be interpreted as a characteristic of the current rule specification rather than as an inherently desirable outcome.

The CLIP and Qwen2.5-VL findings should remain task- and protocol-specific. Prior research has demonstrated the potential of multimodal models for persona-based urban perception and comparison with human judgments (Beneduce et al., 2025; Mushkani, 2026). Other studies have explored interpretable urban concepts and safety-critical transportation reasoning (Yu et al., 2025; Wang et al., 2025). The weaker results of the fixed zero-shot and few-shot configurations evaluated here therefore do not establish a general limitation of vision–language models. Instead, they indicate that this specific three-level road-infrastructure-safety task may require domain adaptation, stronger alignment, or more specialized prompting. CARSA is consequently positioned as a complementary, interpretable baseline and as a potential structured component of future hybrid systems.

Beyond the framework itself, the released benchmark provides a reproducible resource containing 315 contextually categorized street-level images, 945 independent expert ratings, consensus labels, semantic metadata, and evaluation outputs. Together, these resources support further research on perceived road safety, explainable transportation intelligence, and human–AI agreement. The findings nevertheless characterize performance only within the current Mapillary Vistas-derived benchmark and should not be interpreted as evidence of cross-dataset or cross-geographic generalization.

6 Limitations and future work

The current benchmark comprises 315 street-level images and 945 independent assessments from three experts. It provides an initial reproducible resource rather than a comprehensive representation of all geographic, professional, and transportation contexts. Future extensions should therefore include larger expert panels, additional geographic regions, and more diverse road environments. Human-in-the-loop and AI-assisted annotation may support this expansion while preserving annotation quality and explicit inter-rater reliability analysis.

The current study was designed as a paired, label-independent human–framework evaluation on a newly constructed expert-annotated benchmark. The same road scenes necessarily served as common evaluation units for CARSA and the expert panel, whereas the expert-derived perceived-safety labels were created only after the framework specification and predictions had been frozen. Therefore, the current evaluation does not involve reuse of the human Ground Truth for framework development, parameter selection, or calibration.

The reported performance metrics quantify human–framework agreement on the current Mapillary Vistas-derived benchmark and should not be interpreted as evidence of cross-dataset, cross-geographic, or prospective generalization. The predefined parameter-sensitivity analysis evaluates the local robustness of the frozen CARSA specification but does not replace independent external evaluation. A compatible external evaluation would require new images with the required semantic masks and road-context categories, together with a new independently collected expert Ground Truth for Low-, Medium-, and High-safety perception. Prospective evaluation of the unchanged CARSA specification on new datasets, geographic regions, and expert panels is therefore reserved for future work.

Accordingly, the reported results characterize CARSA under reliable semantic inputs rather than an end-to-end image-to-safety pipeline. In a future deployment, automatically predicted masks could introduce upstream perception errors that propagate to the downstream assessment. Because CARSA’s defined inputs are semantic presence and coverage, directly deleting or altering a safety-relevant class in a mask would change the semantic input supplied to CARSA; it would not constitute image-level noise at the contextual reasoning stage. A realistic predicted-mask evaluation would therefore require a specified segmentation model, explicit taxonomy alignment, an empirically grounded error profile, and separate reporting of upstream segmentation errors and downstream CARSA reasoning errors. Such end-to-end system-level validation is reserved for future work.

The current evaluation also assumes correct road-context assignment. Because the assigned context determines both the factor weights and decision thresholds, an incorrect context label could alter the final CARSA score and categorical prediction. Robustness to automatically predicted or uncertain context labels was not quantified in this study and should be examined separately in future end-to-end evaluations.

In a future end-to-end deployment, an explicit input-validity mechanism could abstain when a scene does not provide sufficient observable evidence for reliable assessment. Such abstention would represent an out-of-scope or insufficient-evidence decision rather than a fourth perceived-safety category.

The current framework relies on static imagery and therefore does not directly capture dynamic factors such as traffic density, vehicle speed, pedestrian movement, weather, and changing illumination. Incorporating temporal, multimodal, and real-time data could provide an even richer and more realistic representation of transportation environments.

Human perceptions of safety may also reflect factors beyond visible infrastructure, including familiarity with an area, perceived driver behavior, neighborhood characteristics, and personal security. These influences may help explain some differences between framework predictions and expert judgments, particularly in residential settings. Future work can therefore enrich the contextual representation by combining physical infrastructure with broader perceptual and environmental information.

CARSA was intentionally developed as an interpretable, rule-based framework that prioritizes transparency and explainability. While some assessment decisions therefore depend on expert-informed assumptions and predefined rules, this design makes the reasoning process clear and auditable. Future research may explore hybrid approaches that combine this contextual reasoning with data-driven calibration while preserving interpretability.

The study considered three broad environments including urban, residential, and rural, which capture major contextual differences. Expanding the taxonomy to include highways, school zones, commercial districts, and complex intersections could further improve the framework’s sensitivity and generalizability.

The comparative baseline evaluation was intentionally limited in scope. The CLIP experiment used a single ViT-B/32 backbone and fixed context-aware prompts, while the VLM evaluation considered one Qwen2.5-VL model under zero-shot and nine-example few-shot prompting. The few-shot setting used in-context demonstrations without parameter updating or fine-tuning. Consequently, the reported results should not be generalized to all vision–language models or prompting strategies. Future work may evaluate additional foundation models, larger prompt ensembles, domain-specific adaptation, supervised calibration, and human-alignment mechanisms on independent external benchmarks.

7 Conclusion

This study introduced CARSA, an interpretable and context-aware framework for assessing perceived road-infrastructure safety from street-level imagery. The framework combines semantic presence and coverage, context-specific factor weights, and predefined decision thresholds within a fully specified rule-based procedure. Its development and prediction stages were completed before the independent expert annotation process, preserving the informational separation between framework construction and evaluation.

On the current Mapillary Vistas-derived benchmark, CARSA achieved 92.38% accuracy and substantially outperformed the non-contextual semantic ablation, which achieved 75.87% accuracy. Most disagreements were restricted to adjacent safety categories, while Urban and Rural scenes achieved stronger agreement than Residential scenes. The parameter-sensitivity analysis further showed that the framework was generally stable under moderate factor-weight and normalization changes, while identifying the presence–coverage balance and the Urban Medium–High threshold as influential components.

The released images, expert annotations, semantic metadata, rules, predictions, and evaluation resources provide a reproducible foundation for future research on explainable transportation intelligence and human–AI agreement. However, the reported performance is benchmark-specific and does not establish cross-dataset, geographic, or prospective generalization. Future work should evaluate the unchanged framework on independent expert-annotated datasets, examine predicted semantic masks and uncertain context assignments, and explore hybrid approaches that combine transparent contextual reasoning with data-driven adaptation.

References

• 1

  AlsamhiS. H.KumarS.HawbaniA.ShvetsovA. V.ZhaoL.GuizaniM. (2024). Synergy of human-centered ai and cyber-physical-social systems for enhanced cognitive situation awareness: applications, challenges and opportunities. Cogn. Comput.16 (6), 2735–2755. 10.1007/s12559-024-10271-7

  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  AnkareddyR.DelhibabuR. (2025). Dense segmentation techniques using deep learning for urban scene parsing: a review. IEEE Access13, 34496–34517. 10.1109/ACCESS.2025.3543944

  • CrossRef
  • Google Scholar

• 3

  BeneduceC.LepriB.LucaM. (2025). Urban Safety Perception Through the Lens of Large Multimodal Models: A Persona-based Approach. 10.48550/arXiv.2503.00610

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  BorsosA.CafisoS.Di GrazianoA.PappalardoG. (2026). Transferability of safety inspection procedures for network-wide safety assessment of two-lane rural roads: an Italian-Hungarian experiment. Traffic Inj. Prev.27 (4), 446–454. 10.1080/15389588.2025.2510572

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  CaiQ.Abdel-AtyM.ZhengO.WuY. (2022). Applying machine learning and google street view to explore effects of drivers’ visual environment on traffic safety. Transp. Res. Part C Emerg. Technol.135, 103541. 10.1016/j.trc.2021.103541

  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  CuiQ.ZhangY.YangG.HuangY.ChenY. (2023). Analysing gender differences in the perceived safety from street view imagery. Int. J. Appl. Earth Observation Geoinformation124, 103537. 10.1016/j.jag.2023.103537

  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  DolatyabiP.ReganJ.KhodayarM. (2025). Deep learning for traffic scene understanding: a review. IEEE Access13, 13187–13237. 10.1109/ACCESS.2025.3529289

  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  GraefeJ.PadenS.EngelhardtD.BenglerK. (2023). Human-centered explainability for intelligent vehicles—a user study. Int. J. Human–Computer Interact.39 (16), 3237–3253. 10.1080/10447318.2023.2215098

  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  GuoW.TengZ.SongX.CaoD.CaoH.ZhangW.et al (2024). Toward consumer acceptance of cooperative driving systems: a human-centered shared steering control approach within a hierarchical framework. IEEE Trans. Consumer Electron.70 (1), 635–645. 10.1109/TCE.2024.3357985

  • CrossRef
  • Google Scholar

• 10

  HamimO. F.UkkusuriS. V. (2024). Towards safer streets: a framework for unveiling pedestrians’ perceived road safety using street view imagery. Accid. Analysis and Prev.195, 107400. 10.1016/j.aap.2023.107400

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 11

  HamimO. F.UkkusuriS. V. (2022). Determining prominent factors across system hierarchies to improve road safety in lmics: a case study of Bangladesh. Saf. Sci.150, 105709. 10.1016/j.ssci.2022.105709

  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  HouX.ZhaoY. (2024). Analysis of road safety perception and influencing factors using street view imagery and deep learning. ISPRS Int. J. Geo-Information13 (8), 272. 10.3390/ijgi13080272

  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  HuangJ.QingL.HanL.LiaoJ.GuoL.PengY. (2023). A collaborative perception method of human–urban environment based on machine learning and its application to the case area. Eng. Appl. Artif. Intell.119, 105746. 10.1016/j.engappai.2022.105746

  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  HussainS. M. F.HamzaS. M.SamadA. (2022). “Image segmentation for autonomous driving using u-net inception,” in 2022 7th International Conference on Signal and Image Processing (ICSIP), 426–429. 10.1109/ICSIP55141.2022.9885809

  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  JoyceR. J.AmlaniD.NicholasC.RaffE. (2023). Motif: a malware reference dataset with ground truth family labels. Comput. and Secur.124, 102921. 10.1016/j.cose.2022.102921

  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  KabirI.RezaM. A.BillahS. (2025). “Logic-rag: augmenting large multimodal models with visual-spatial knowledge for road scene understanding,” in 2025 IEEE International Conference on Robotics and Automation (ICRA), 10585–10591. 10.1109/ICRA55743.2025.11128818

  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  KangY.ChenJ.LiuL.SharmaK.MazzarelloM.MoraS.et al (2026). Decoding human safety perception with eye-tracking systems, street view images, and explainable ai. Comput. Environ. Urban Syst.123, 102356. 10.1016/j.compenvurbsys.2025.102356

  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  KhairnarS.ThepadeS. D.KolekarS.GiteS.PradhanB.AlamriA.et al (2025). Enhancing semantic segmentation for autonomous vehicle scene understanding in indian context using modified canet model. MethodsX14, 103131. 10.1016/j.mex.2024.103131

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  KonstantinidisM.LeL. W.GaoX. (2022). An empirical comparative assessment of inter-rater agreement of binary outcomes and multiple raters. Symmetry14 (2), 262. 10.3390/sym14020262

  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  KuznietsovA.GyevnarB.WangC.PetersS.AlbrechtS. V. (2024). Explainable ai for safe and trustworthy autonomous driving: a systematic review. IEEE Trans. Intelligent Transp. Syst.25 (12), 19342–19364. 10.1109/TITS.2024.3474469

  • CrossRef
  • Google Scholar

• 21

  LiJ.CheangC. F.YuX.TangS.DuZ.ChengQ. (2025a). A segmentation network for enhancing autonomous driving scene understanding using skip connection and adaptive weighting. Sci. Rep.15 (1), 36692. 10.1038/s41598-025-20592-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  LiQ.LiY.ZhangS.ZhouX.PanZ. (2025b). A theoretical framework for human-centered intelligent information services: a systematic review. Inf. Process. and Manag.62 (1), 103891. 10.1016/j.ipm.2024.103891

  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  LiuY.ChenT.ChungH.JangK.XuP. (2025). Is there an emotional dimension to road safety? A spatial analysis for traffic crashes considering streetscape perception and built environment. Anal. Methods Accid. Res.46, 100374. 10.1016/j.amar.2025.100374

  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  LuC.LiJ.ChenT.ZhouC.FanY.RenX.et al (2026). From algorithm to expert: Rlhf-guided vision-language model for 3d-eem fluorescence spectroscopy matching. Comput. Mater. and Continua87 (2), 1–10. 10.32604/cmc.2026.075400

  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  MuhammadK.HussainT.UllahH.SerJ. D.RezaeiM.KumarN.et al (2022). Vision-based semantic segmentation in scene understanding for autonomous driving: recent achievements, challenges, and outlooks. IEEE Trans. Intelligent Transp. Syst.23 (12), 22694–22715. 10.1109/TITS.2022.3207665

  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  MushkaniR. (2026). Benchmarks for vision–language models in urban perception should be reliability-aware and negotiated. To Appear Proc. 43rd Int. Conf. Mach. Learn. (ICML 2026). 10.48550/arXiv.2606.00871

  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  QinT.WegheN. V.HuangH. (2025). The role of urban built environments in the perception of safety: a systematic review. Environ. Impact Assess. Rev.114, 107915. 10.1016/j.eiar.2025.107915

  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  RandhawaS.Ayg”unE.RandhawaG.HerfortB.LautenbachS.ZipfA. (2025). Paved or unpaved? A deep learning derived road surface global dataset from mapillary street-view imagery. ISPRS J. Photogrammetry Remote Sens.223, 362–374. 10.1016/j.isprsjprs.2025.02.020

  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  RidleyM. (2025). Human-centered explainable artificial intelligence: an annual review of information science and technology (arist) paper. J. Assoc. Inf. Sci. Technol.76 (1), 98–120. 10.1002/asi.24889

  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  RitaL.NathvaniR.PeliteiroM.BostanT.-C.MullerE.SuelE.et al (2023). Using deep learning and google street view imagery to assess and improve cyclist safety in london. Sustainability15 (13), 10270. 10.3390/su151310270

  • CrossRef
  • Google Scholar
  • View reference in article

• 31

  SaitoY.OgumaY.InoueS.BreugelmansR.KikuchiH.OkaK.et al (2022). Inter-rater reliability of streetscape audits using online observations: microscale audit of pedestrian streetscapes (maps) global in Japan. Prev. Med. Rep.30, 102043. 10.1016/j.pmedr.2022.102043

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 32

  SchrankA.WalochaF.BrandenburgS.OehlM. (2024). Human-centered design and evaluation of a workplace for the remote assistance of highly automated vehicles. Cognition, Technol. and Work26 (1), 183–206. 10.1007/s10111-024-00753-x

  • CrossRef
  • Google Scholar

• 33

  SiebkeC.MaiM.ProkopG. (2023). What do traffic simulations have to provide for virtual road safety assessment? Human error modeling in traffic simulations. IEEE Trans. Intelligent Transp. Syst.24 (2), 1419–1436. 10.1109/TITS.2022.3220961

  • CrossRef
  • Google Scholar

• 34

  SunQ.LiY.AlturkiE.MurthyS. M. K.SchullerB. W. (2026). Towards friendly ai: a comprehensive review and new perspectives on human-ai alignment. AI Ethics6, 193. 10.1007/s43681-026-01070-x

  • CrossRef
  • Google Scholar

• 35

  TbailehI.BagriyanikS. (2025). Visual quality assessment of e-commerce product images using convolutional neural networks. Multimed. Syst.31, 428. 10.1007/s00530-025-02009-8

  • CrossRef
  • Google Scholar

• 36

  TbailehI.RadwanA.YaseenO.AshqarH. I.ElhenawyM. (2026). “Roadvision: a comprehensive dataset for object detection in real-world road environments,” in Intelligent Systems, Blockchain, and Communication Technologies. Editors AbdelgawadA.HameedA. A.JamilA. (Cham: Springer), 1801, 1–15. 10.1007/978-3-032-15784-322

  • CrossRef
  • Google Scholar

• 37

  WangR.HuangC.YeY. (2024). Measuring street quality: a human-centered exploration based on multi-sourced data and classical urban design theories. Buildings14 (11), 3332. 10.3390/buildings14113332

  • CrossRef
  • Google Scholar

• 38

  WangM.ChenY.PangA.CaiY.ChenC. S.KanY.et al (2025). “VLMLight: Safety-critical traffic signal control via vision–language meta-control and dual-branch reasoning architecture,” Adv. Neural Inf. Process. Syst., 38, 44266–44297. 10.52202/085713-1320

  • CrossRef
  • Google Scholar

• 39

  WilleminkM. J.MastrodicasaD.MadaniM. H.CodariM.ChepelevL. L.MistelbauerG.et al (2023). Inter-observer variability of expert-derived morphologic risk predictors in aortic dissection. Eur. Radiol.33, 1102–1111. 10.1007/s00330-022-09056-z

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 40

  WuS.WangW.ZhangC.ZengQ.WangH.LinJ.et al (2026). Exploring multimodal large language models’ potential in simulating human perception of urban cycling environments: a street-view perspective. Inf. Geogr.2 (1), 100037. 10.1016/j.infgeo.2025.100037

  • CrossRef
  • Google Scholar

• 41

  YangZ.ZhaoC.MaedaH.SekimotoY. (2022). Development of a large-scale roadside facility detection model based on the mapillary dataset. Sensors22 (24), 9992. 10.3390/s22249992

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 42

  YangW.WeiY.WeiH.ChenY.HuangG.LiX.et al (2023). Survey on explainable ai: from approaches, limitations and applications aspects. Human-Centric Intell. Syst.3 (3), 161–188. 10.1007/s44230-023-00038-y

  • CrossRef
  • Google Scholar

• 43

  YapW.BiljeckiF. (2023). A global feature-rich network dataset of cities and dashboard for comprehensive urban analyses. Sci. Data10, 667. 10.1038/s41597-023-02578-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 44

  YeY.ZhongC.SuelE. (2024). Unpacking the perceived cycling safety of road environment using street view imagery and cycle accident data. Accid. Analysis and Prev.205, 107677. 10.1016/j.aap.2024.107677

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 45

  YeX.DuJ.LiX.ShawS. L.FuY.DongX.et al (2025a). Human-centered geoai foundation models: where geoai meets human dynamics. Urban Inf.4 (1), 2. 10.1007/s44212-025-00067-x

  • CrossRef
  • Google Scholar

• 46

  YeX.LiS.GongW.LiX.LiX.DadashovaB.et al (2025b). Street view imagery in traffic crash and road safety analysis: a review. Appl. Spatial Analysis Policy18, 50. 10.1007/s12061-025-09653-7

  • CrossRef
  • Google Scholar

• 47

  YinM.SingM. C. P.YungE. H. K. (2026). Can chatgpt-5 assess human perceptions of dynamic urban streetscape transitions? Insights from Hong Kong. Build. Environ.290, 114155. 10.1016/j.buildenv.2025.114155

  • CrossRef
  • Google Scholar

• 48

  YuX.MaJ.TangY.YangT.JiangF. (2024). Can we trust our eyes? Interpreting the misperception of road safety from street view images and deep learning. Accid. Analysis and Prev.197, 107455. 10.1016/j.aap.2023.107455

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 49

  YuY.YuZ.ShiX.WanR.WangB.ZhangJ. (2025). Towards transparent urban perception: a concept-driven framework with visual foundation models. ISPRS Int. J. Geo-Information14 (8), 315. 10.3390/ijgi14080315

  • CrossRef
  • Google Scholar

• 50

  YuanK.LanH.GaoY.WangK. (2026). Interpretable multimodal framework for human-centered street assessment: integrating visual-language models for perceptual urban diagnostics. Land15 (3), 449. 10.3390/land15030449

  • CrossRef
  • Google Scholar