Design of a vehicle navigation system integrating CART decision tree and DBSCAN algorithm

Abstract

Introduction:

Accurate road type recognition and pavement information aggregation are essential for reliable vehicle navigation in complex urban conditions.

Methods:

This paper integrates a CART decision tree with DBSCAN clustering within an edge-cloud architecture. CART classifies roads using 28-dimensional sensor features via Gini gain optimization and post-pruning. DBSCAN (Eps = 30 m, MinPts = 3) aggregates and denoises pavement information points using Haversine distance.

Results:

CART achieved 96.8% accuracy, 97.2% precision, 97.3% recall, and a 97.2% F1 score, outperforming RF, SVM, and LR with minimal model size and inference time. DBSCAN merged 85.7% of redundant points with noise below 5%. Real-vehicle tests showed 98% + recognition accuracy for non-ordinary urban roads, a 92.5% detection rate, 89.2% warning success rate, and a 126 m average advance warning distance.

Discussion:

The CART-DBSCAN integration provides an effective closed-loop navigation solution from perception to warning. Future work will incorporate multi-modal sensors and domain adaptation to enhance robustness across diverse scenarios.

1 Introduction

As urban traffic becomes increasingly complex, vehicle navigation systems have expanded from path planning to driving assistance and safety warnings (Nandhu et al., 2026). Accurately identifying road types, locating curves, and road anomalies has become the key to improving navigation reliability. However, the motion characteristics of intersection turns and continuous curves are similar, and the traditional threshold method is prone to misjudgment. At the same time, data collected by multiple terminals has redundancy and spatial offset due to positioning errors and noise interference, which directly affects the accuracy of early warning. In recent years, domestic and foreign scholars have conducted extensive research on road identification and information fusion issues in vehicle navigation systems. Chen et al. proposed the GoComfort framework to solve the problem of low accuracy in traditional road damage detection. This framework used edge-cloud hybrid computing paradigm to collect high-precision road damage-related data, and designed an accurate road damage recognition model that integrates point cloud spatial structure and street view image texture features. The results showed that the road damage identification accuracy of this method reached 87.5%, and the user comfort acceptance rate reached 93.8% (Chen et al., 2022). Lei et al. proposed a multi-objective autonomous driving navigation method embedded in digital twins to solve the problem of low efficiency caused by traditional path planning ignoring road conditions. Experimental results showed that this method showed good effectiveness and robustness in a variety of construction scenarios, and significantly improved system performance and reliability (Lei et al., 2024). Zhao et al. proposed an enhanced direction method that combined distributed inertial navigation systems, polarization, and geomagnetic compasses, and introduced variational Bayesian federated volumetric Kalman filtering and residual-based distributed monitoring compensation algorithms to solve the problem of insufficient accuracy and robustness of multi-source fusion navigation systems in complex environments. The results showed that the system could accurately estimate and dynamically adjust the noise covariance based on the measurement quality, effectively reduced the impact of occlusion and electromagnetic interference, and enhanced the accuracy and robustness of the navigation system (Zhao et al., 2024). Tang et al. proposed a vehicle heterogeneous multi-source information fusion positioning method based on information probability. By unifying heterogeneous navigation sources into an information probability model and achieving rapid fusion based on relative entropy, it solved the problems of insufficient accuracy of a single navigation source and low generalization ability of multi-source fusion methods. The results showed that this method had high positioning accuracy, strong real-time performance, and can effectively suppress interference (Tang et al., 2024).

In terms of road type discrimination, Liu et al. Proposed a road type recognition method based on fusion features to solve the problem of complex road conditions of unmanned tracked vehicles, combining the depth features of the image with the statistical features of the vertical acceleration of the mount mass, and using the machine learning classification algorithm to realize the road type recognition. The experimental results show that the overall accuracy of support vector machine (SVM) and random forest (RF) can reach more than 90% (Liu J. et al., 2023). Wei et al. Proposed a two-stage road perception framework, which obtained parameters with strong generalization ability through pre training on large-scale labeled data sets, and then fine tuned on small sample labeled data sets with the help of mask attention mechanism. The weighted joint loss function was used to solve the problem of category imbalance, realizing the accurate identification of road types and the adaptive evaluation of pavement roughness (Wei et al., 2024).

The existing road type discrimination research has made a series of progress in feature extraction and model design, but there are still shortcomings. On the one hand, most methods rely on manually designed statistical features or image features, which have limited robustness to sensor noise and vehicle dynamic coupling interference, and the recognition accuracy is prone to fluctuate under the urban conditions of frequent start and stop and complex steering. On the other hand, the generalization ability of existing models is mostly verified only on specific acquisition vehicles and fixed routes, lacking systematic evaluation on cross time and cross road conditions. In addition, the existing research generally takes the road type recognition results as the final output, but does not form a linkage with the downstream road information spatial aggregation link, which is difficult to support the complete closed-loop navigation application from perception to early warning. In view of the above shortcomings, the cart decision tree is introduced to build the road type recognition model. Through the Gini gain maximization splitting strategy and post pruning mechanism, the classification accuracy and generalization ability are improved while maintaining the lightweight of the model. At the same time, the recognition results are used as the pre input of DBSCAN clustering, and the integration of road type discrimination and road information aggregation is realized through the end cloud collaborative architecture, which is clearly different from the existing research.

The innovation points of the research are: (1) Applying the CART decision tree to the road type two-classification task, and achieving high-precision identification under complex working conditions through windowed multi-dimensional feature extraction and post-pruning strategies. (2) Using DBSCAN to achieve adaptive aggregation of precise road information, and adapting to the longitude and latitude coordinate space through the Haversine distance measurement. (3) Building a edge-cloud collaborative system architecture to achieve integration of road identification and information clustering, and provide an engineering implementation technical solution for intelligent vehicle navigation systems.

2 Methods and materials

2.1 Road type identification method based on CART decision tree

Accurately distinguishing ordinary urban roads from non-ordinary urban roads is the key to ensuring the reliability of vehicle navigation systems. Turning at intersections on ordinary urban roads is often accompanied by complex working conditions such as low-speed driving and starting and stopping of traffic lights. If it is confused with continuous curves on non-ordinary urban roads, the accuracy of curve curvature identification will be seriously affected. To this end, this study proposes a road type identification method based on CART decision tree, which uses multi-dimensional sensor data collected during vehicle driving to make a binary classification judgment on the current road type. The core of cart decision tree is to divide the feature space into several regions by recursive dichotomy, so that the samples in each region belong to the same category as much as possible. In the task of road type recognition, the algorithm takes the windowed 28 dimensional statistical features as the input and the road type label as the output. During the training, the algorithm starts from the root node, traverses all features and all possible bisection points, and selects the feature and bisection point that maximize the Gini gain as the splitting rule of the current node. Then recursively execute the above process on the divided two child nodes until the preset maximum depth is reached or the number of node samples is lower than the threshold. During the prediction, the new sample vector starts from the root node, judges down layer by layer according to the splitting conditions of each node, and finally falls into a leaf node. The category with the highest proportion in the leaf node is the prediction result. By limiting the depth of the tree through the post pruning strategy, the model can effectively avoid over fitting the noise in the training data, so as to improve the generalization ability of the model on the unknown road data. The CART decision tree is in Figure 1.

FIGURE 1

From Figure 1, the decision tree is a classification model based on a tree structure, in which the CART uses the Gini coefficient as the evaluation index for node splitting (Chen, 2025; Dai and Li, 2024). Among them, the calculation of the Gini coefficient can be expressed as Formula 1.

In Formula 1, is a given sample set, is the category, and represents the proportion of samples in the th category. In addition, the CART decision tree achieves classification by recursively dividing the feature space, and each division selects the features and split points that reduce the Gini coefficient the most (Rafika et al., 2025). The divided Gini coefficient can be expressed in Formula 2.

In Formula 2, represents the feature used to divide the sample set, and and are the two subsets obtained after dividing according to the feature . represents the weighted Gini coefficient divided by feature (Yang et al., 2025). The feature that minimizes the Gini coefficient after division is selected as the split node, that is, the Gini gain is maximized in Formula 3.

In Formula 3, represents the decrease in the Gini coefficient after dividing by feature , that is, the Gini gain. Before applying the CART decision tree for road type identification, it is necessary to construct a sample data set that can characterize the road characteristics. The specific process is in Figure 2.

FIGURE 2

The total sample size of the dataset constructed for the study is 14,280 time windows, with each window containing 1000 sampling points. Randomly divided into a training set and a testing set in an 8:2 ratio, with the training set containing 11,424 samples and the testing set containing 2856 samples. The distribution of two types of road labels is 9820 for ordinary urban roads and 4460 for non ordinary urban roads, with a certain degree of category imbalance but within an acceptable range. The data collection was completed on 5 urban roads and 3 expressways in Chengdu, covering a total distance of about 65 km. The collection period covers four time periods on weekdays: morning peak (7:00–9:00), noon off peak (11:00–14:00), evening peak (17:00–19:00), and early morning (0:00–5:00). The weather conditions include sunny, cloudy, and light rain. The 28 dimensional feature vector consists of 4 dimensions including the median, mean, standard deviation, and maximum value of the vehicle speed, 12 dimensions including 4 dimensions for each of the three-axis accelerations, and 12 dimensions including 4 dimensions for each of the three-axis gyroscopes, for a total of 28 dimensions. When annotating, take the average latitude and longitude of all sampling points within each time window, call the Gaode inverse geographic coding interface to obtain the road name at that location, and then map the road name to a binary label according to the Chengdu road classification standard. Urban main roads, secondary roads, and branch roads are classified as ordinary urban roads, while urban expressways, highways, and ramps are classified as non ordinary urban roads. The mapping rules are manually reviewed to ensure accuracy. After the sample construction is completed, the CART is used to train the decision tree with the goal of maximizing the Gini gain. To avoid model overfitting, the study adopts the method of limiting the maximum depth of the decision tree for post-pruning, determines the optimal depth parameters through grid search, and finally realizes the identification of road types.

2.2 Accurate pavement information clustering method based on DBSCAN

After completing the road type identification, the system can filter out the curve information in non-ordinary urban roads and form preliminary accurate road information data points. However, due to the use of group sensing mode, when multiple collection terminals identify the same road anomaly or curve, affected by factors such as positioning accuracy and sensor noise, multiple data points with similar longitude and latitude but not completely overlapping will be generated. If these data points are directly stored in the road surface accurate information database, it will not only cause data redundancy, but also cause information offset in subsequent navigation applications due to positioning errors. Therefore, it is necessary to perform clustering processing on the initially identified data points, merge multiple records describing the same actual location into one representative data point, and at the same time filter out outlier noise points. To this end, the study proposes a density clustering method based on DBSCAN, which is used to cluster precise pavement information, as shown in Figure 3.

FIGURE 3

DBSCAN algorithm does not preset the number of clusters, but automatically finds clusters of arbitrary shape through the density connectivity in the neighborhood of sample points. For the task of precise pavement information clustering, the algorithm takes all data points in each class as input, and retrieves all adjacent points within the 30 m neighborhood of each point that has not been accessed. If the number of adjacent points in the neighborhood of the point reaches minpts = 3, the point is marked as the core point, and a new cluster is created with the point as the core. At the same time, all unreachable points in the neighborhood are accessed recursively, and the points that meet the density conditions are included in the current cluster in turn. If the number of adjacent points is less than minpts but is located in the neighborhood of a core point, it is marked as a boundary point and classified into the corresponding cluster. If a point is neither a core point nor in the neighborhood of any core point, it is determined as a noise point and eliminated. The longitude and latitude coordinates of all points in the cluster are combined into a representative point by arithmetic mean. The above process is performed independently on eight class subsets, so as to realize the effective combination of multi-source acquisition points at the same location and the filtering of outlier noise points. In accurate pavement information clustering, the distance between sample points is calculated using the spherical distance formula to adapt to the longitude and latitude coordinates of the earth’s surface. It assumes that the latitudes of two points and are and , respectively, and the longitudes are and . The spherical distance between the two points can be calculated by the Haversine formula in Formula 4.

In Formula 4, is the average radius of the earth, which is 6371 km. For the evaluation of clustering results, the silhouette coefficient is used to measure the clustering quality in Formula 5.

In Formula 5, represents the average distance from sample point to other sample points in the same cluster, and represents the average distance from sample point to all sample points in the nearest other clusters (Lian et al., 2023). The clustering operation process can be seen in Figure 4.

FIGURE 4

From Figure 4, the research first extracts all data points from the preliminary road surface precise information database, and performs clustering processing according to direction attributes and types. The direction attribute is divided into four digits, which represent different combinations of north-south and east-west directions (Niu et al., 2022). The types include road surface anomalies and curve curvature. A total of eight categories are clustered separately. For each type of data point, the values of neighborhood radius Eps and minimum neighborhood sample size MinPts are jointly determined through grid search and contour coefficient evaluation. The candidate range for Eps is 10 m–60 m with an interval of 5 m, while the candidate range for MinPts is 2–8 with an interval of 1. Perform clustering on eight subsets of data for each parameter combination and calculate the average silhouette coefficient. Select the parameter combination with the highest silhouette coefficient as the optimal configuration. The initial selection range of Eps also refers to the k-distance graph of each data subset, and the position where the inflection point appears in the graph is used as the reasonable lower bound of the neighborhood radius. The selection of MinPts is based on the minimum redundancy requirement for at least 3 different collection terminals to generate reporting records for the same location. After the algorithm is executed, multiple data points in the same cluster are merged into one representative data point by calculating the mean of longitude and latitude, and the noise points are directly eliminated. After the clustering is completed, the longitude and latitude coordinates of the representative data points are converted from the WGS84 coordinate system to the GCJ02 coordinate system to adapt to the Gaode map service, and are finally stored in the final road surface accurate information database.

2.3 Design of vehicle navigation system integrating CART and DBSCAN

The road type identification method based on CART decision tree and the precise road information clustering method based on DBSCAN respectively solve the two problems of road type identification and data point aggregation. However, in a complete vehicle navigation system, road type identification and road surface information clustering do not exist in isolation. The two need to be organically combined to form a complete link from data collection to information application. To this end, the study integrates the two methods to design a complete vehicle navigation system architecture to realize the identification, processing and navigation application of precise road information, as shown in Figure 5.

FIGURE 5

From Figure 5, the system as a whole adopts the working mode of terminal-cloud collaboration and consists of three parts: data collection terminal, cloud processing platform and navigation application terminal. The data collection terminal includes vehicle Controller Area Network (CAN) bus and Android mobile phone collection App. It is responsible for collecting vehicle speed, three-axis acceleration, three-axis gyroscope, longitude and latitude and other status data during vehicle operation, and reports it to the Alibaba Cloud IoT platform in real time through the Message Queuing Telemetry Transport (MQTT) protocol. After receiving the reported data, the cloud processing platform first stores the original data in the Alibaba Cloud Relational Database Service (RDS) database, and then triggers the data processing process (Liu W. et al., 2023; Mo et al., 2022). In the data processing stage, the system first performs Kalman filtering on the raw sensor data to eliminate noise interference caused by the inherent vibration of the road and the vibration transmitted by the engine. The basic form of Kalman filter can be expressed as Formula 6.

In Formula 6, is the state prior estimate, is the state posterior estimate, is the state transition matrix, is the control input matrix, is the control vector, is the observation matrix, and and are the process noise covariance and observation noise covariance, respectively. is the prior estimated error covariance. is the Kalman gain, and is the actual observation vector. In this system, the signal to be estimated by Kalman filter is the real motion state of the vehicle, including the real values of longitudinal speed, lateral speed, yaw rate and three-axis acceleration. The observed values are from the vehicle speed and yaw rate output by the vehicle can bus, and the three-axis acceleration and three-axis angular velocity output by the mobile inertial measurement unit. The filter parameters are calibrated off-line according to the vehicle kinematics model and the statistical characteristics of sensor noise. The state transition matrix and observation matrix are determined by the constant speed model. The process noise covariance and observation noise covariance are obtained through the analysis of sensor data under static and constant speed cruise conditions, respectively. The above parameters are fixed after calibration.

The filtered data is sliced according to time windows, and each window contains 1000 sampling points. The median, mean, standard deviation and maximum value of each sensor data within the window are extracted as feature vectors (Wang et al., 2023). This feature vector is input into the trained CART decision tree model to determine which type of road the current vehicle is on. If the determination result is a non-ordinary urban road, the curve curvature is further calculated based on the vehicle speed and yaw angular velocity in Formula 7.

In Formula 7, is the road curvature, and is the vehicle yaw angular velocity, is the vehicle running speed. The system takes the feature vector in each 10s time window as the input, and the cart decision tree starts from the root node, and then makes binary judgment according to the splitting characteristics and segmentation threshold stored in each node until it reaches the leaf node, and outputs the road type label corresponding to the current window. If the output is an ordinary urban road, all sampling points in the current window will not trigger subsequent processing. If the output is a non ordinary urban road, the curvature value of each sampling point in the window is calculated according to Formula 7 and compared with the preset curvature threshold. The sampling point whose curvature value exceeds the threshold is marked as a valid detection point, and its spatial longitude, latitude and direction characteristics are retained. The above marked points are aggregated to form a preliminary pavement accurate information data set. The data points in the preliminary pavement accurate information database then enter the DBSCAN. The system groups data points according to their type and direction attributes, and performs DBSCAN on each group of data points (Zhang et al., 2022; Bajal et al., 2022). After clustering is completed, the data points in the same cluster are merged into a representative data point. Its longitude and latitude are taken as the mean of all points in the cluster, the curvature value is taken as the maximum value of all points in the cluster, and the direction attribute retains the direction of most points in the cluster (Yang et al., 2022; Yang et al., 2023). All location points that have been clustered and merged will be integrated into the final version of the precise pavement information data set after geographic coordinate conversion. The details of the navigation application terminal can be seen in Figure 6.

FIGURE 6

From Figure 6, at the navigation application level, users set the starting point and end point through the Android mobile phone safety navigation App. The App calls the Gaode map SDK to obtain the planned route and uploads the route information to the cloud application service. The application service calls the data point matching algorithm, retrieves the data points located on the planned path in the final road surface accurate information database, and filters them based on the matching degree between the vehicle driving direction and the direction attribute of the data point (Pang et al., 2024; Chen, 2022). During the matching process, the ray method is used to determine whether the data point is located within the rectangular area constructed by the path. The successfully matched data points are sorted in the order in which the vehicle passes and then sent to the App. The App calculates the distance between the current position and the nearest data point in real time, and issues an early warning to the driver when the distance is appropriate. The decision logic of early warning trigger is that the system obtains the real-time longitude and latitude of the vehicle in a period of 1s, calculates the spherical distance between it and all the matching road surface accurate information data points on the planned path one by one, and records the distance between the current vehicle and the nearest target point. When the distance value is less than the preset warning distance threshold for the first time, the system will immediately trigger the warning, and simultaneously send voice prompts and interface pop-up warnings at the navigation app. The threshold of early warning distance is set according to the type of target points. The threshold of sharp turning point is set to 150 m, the threshold of slow turning point is set to 120 m, and the threshold of abnormal points on the road is set to 100 m, respectively corresponding to the safety response distance required by drivers under different road characteristics. After the vehicle passes the target point, the system will automatically clear the warning status of the point and switch to the monitoring of the next target point. Based on the above, the system can realize a closed loop from collection, processing and application of accurate pavement information.

3 Results

3.1 Algorithm performance test

To quantitatively evaluate the core algorithm of the proposed vehicle navigation system, the study separately verified the classification performance of the CART road type recognition model and the effectiveness of the DBSCAN clustering algorithm in merging accurate road surface information on independent test data sets. The test data was collected and constructed in the same way as the training set, and was also derived from data collected through the CAN bus and mobile phone sensors during actual vehicle driving. The hardware environment used in the experiment was: the data collection terminal included a test vehicle equipped with CAN bus (the model was Geely Boyue 2021) and an Android smartphone (Xiaomi 11, system version was Android 12), and the cloud server was configured as an Alibaba Cloud ECS instance (4-core vCPU, 16 GB memory, operating system was Ubuntu 20.04). The software environment was: data processing and algorithm implementation were based on Python 3.8, and the main dependent libraries included scikit-learn 1.0.2, pandas 1.4.0 and numpy 1.21.0. The experiment adopted a comparative method and introduced three classification algorithms: Support Vector Machine (SVM), Random Forest (RF) and Logistic Regression (LR) as baselines. To ensure the fairness of the experiments, all algorithms were trained and tested on the same feature dataset (28-dimensional features), and grid search was used to optimize their hyperparameters. The performance test results of the CART decision tree is in Figure 7.

FIGURE 7

Figure 7a showed that the accuracy of all models increased with the expansion of data size, and the improvement slowed down after the data volume reached 60%. The final accuracy rates of RF, SVM, and LR were 0.961, 0.944, and 0.925, respectively. The CART decision tree steadily improved from 0.832 to 0.968, maintaining the lead under all data sizes. In terms of accuracy in Figure 7b, CART increased from 0.840 to 0.972, always ranking first. RF, SVM, and LR were 0.963, 0.950, and 0.931, respectively, indicating that the proportion of samples predicted by CART to be positive was actually the highest. In terms of recall rate in Figure 7c, CART improved from 0.845 to 0.973, which was the best performance, and CART and RF were better than that of SVM and LR at each data scale, indicating that the tree-based model could more effectively identify target road conditions such as curves. In the F1 score of Figure 7d, RF was 0.965, SVM was 0.950, LR was 0.935, and CART was as high as 0.972, indicating that its comprehensive performance and stability were the best. In addition, experiments evaluated the training efficiency and complexity of the model, and the results are shown in Figure 8. It is worth noting that the model complexity level is divided by the core structural parameters such as the maximum depth of the decision tree, the order of the kernel function of the support vector machine and the number of trees in the random forest. Level 1 corresponds to the model configuration with the simplest structure, level 5 corresponds to the model configuration with the most complex structure, and the intermediate level increases evenly within the range of parameter values. Each model was trained at five levels of complexity and its training time, storage volume and single inference time were recorded to evaluate the impact of different structural complexity on the efficiency of the model.

FIGURE 8

Figure 8a shows the model training time results. As the model complexity increased from level 1 to level 5, the training time of the four models increased significantly. The training time of the CART decision tree remained the lowest throughout the entire process, increasing from 0.5s to 6.2s, with the slowest growth. RF reached 45.1s, which was always much longer than other models. The training time of SVM increased from 2.1s to 22.3s. LR had the shortest training time, second only to CART, and took up to 6.5s. Figure 8b showed the model size results. As the complexity increased, the volumes of the four models expanded dramatically. CART decision trees always had the smallest model size, with a maximum of only 105 KB. RF reached 2200 KB at level 5, SVM reached 4500 KB, and LR reached 135 KB. Figure 8c showed the time-consuming result of model inference. As the complexity increased, the inference time of the model increased simultaneously. The average single inference time of CART decision tree was the shortest in the entire process, reaching a maximum of only 0.32 m. The inference time of RF was the highest, with a maximum of 3.80 m, and SVM was 1.70 m. The inference time of LR was lower, but still higher than that of CART, up to 0.52 m. Taken together, CART decision tree had obvious advantages in the three efficiency dimensions of training, storage and reasoning.

The DBSCAN parameter grid search uses contour coefficient as the evaluation index, and the results are shown in Figure 9.

FIGURE 9

Taking the “bend northeast” subset as an example, its k-distance curve is shown in Figure 9a, with a clear inflection point around 30 m, indicating that this distance is the appropriate threshold lower limit for distinguishing between cluster points and noise points. The average contour coefficient distribution under different parameter combinations is shown in Figure 9b heatmap, where the horizontal axis is Eps and the vertical axis is MinPts. The heatmap shows that when MinPts is fixed at 3, the contour coefficient of Eps reaches its peak and remains stable in the range of 25 m–35 m, with the highest contour coefficient corresponding to 30 m. As MinPts increases to 4 or above, the overall contour coefficient shows a decreasing trend, indicating that excessively high density thresholds can cause homologous data points that should have been merged to be split into multiple clusters. When MinPts is set to 2, the contour coefficient is also lower than the combination of MinPts = 3, indicating that a low density threshold is difficult to effectively distinguish noise points. Based on the comprehensive grid search results and k-distance graph analysis, Eps = 30 m and MinPts = 3 were ultimately determined as the optimal parameter combination.

For the DBSCAN clustering algorithm, it mainly evaluates its ability to spatially merge and denoise the initially identified precise road information data. The experiment used preliminary data points collected by the test vehicle on a specific road section. The road section contained multiple known actual curves and road abnormal points. Each actual point generated an average of 5–8 data points with similar positioning due to repeated collection by multiple vehicles. The experiment tested the clustering effect under different neighborhood radii for three typical sub-datasets: “Curve - NorthEast”, “Curve - SouthWest”, and “Road Anomaly - North-South”, as shown in Table 1.

Table 1 shows the clustering effect of different road feature sub-data sets under their respective adaptation neighborhood radii. Eps = 30 m was the globally optimal parameter, and the results of Eps = 20 m and Eps = 50 m were used to illustrate the performance changes caused by too small or too large radii, respectively, thereby proving the balance and robustness of parameter selection. When Eps was set to 30 m and MinPts was 3, the algorithm could achieve high silhouette coefficients (both greater than 0.65) on different categories of data, indicating high internal tightness of the cluster and good separation between clusters. Under this parameter, the merging rate of data points reached an average of 85.7%, while the elimination rate of noise points was controlled within 5%, effectively eliminating redundant data caused by positioning errors while retaining the vast majority of effective information. Based on the above, the experimental results showed that the CART decision tree model and the DBSCAN algorithm together constituted an efficient and reliable road identification and data processing solution in the vehicle navigation system.

To quantify the independent contributions of each module to the overall performance of the system, ablation experiments were conducted. Four system variants were set up in the experiment: removing Kalman filtering and directly using raw sensor data, replacing DBSCAN clustering with simple mean merging of latitude and longitude, replacing CART decision tree with a simple discrimination rule based on dual thresholds of vehicle speed and yaw rate, and using the complete system as a reference. All variants were evaluated on the same test dataset, with road type recognition accuracy and warning success rate as evaluation metrics. The results are shown in Table 2.

From Table 2, it can be seen that the recognition accuracy of the complete system is 96.8%, which drops to 91.2% without Kalman filtering, indicating that filtering has a significant effect on eliminating sensor noise and stabilizing feature extraction. After using simple mean merging instead of DBSCAN, the success rate of early warning decreased from 89.2% to 76.5%. The reason for this is that the outlier noise points that were not filtered out by density clustering were directly stored in the database, resulting in frequent false triggering of the early warning system. After replacing with simple threshold rules, the recognition accuracy is only 82.3%, far lower than CART decision trees, which verifies the necessity of machine learning models in multi-dimensional feature joint discrimination.

In addition, using the built-in Gini importance evaluation of CART decision tree to assess the contribution of each feature to classification, the importance ranking of 28 dimensional features is shown in Figure 10.

FIGURE 10

As shown in Figure 10, the top five features are the standard deviation of vehicle speed, the mean yaw rate, the standard deviation of lateral acceleration, the mean of vehicle speed, and the standard deviation of vertical acceleration, indicating that the fluctuation index of vehicle motion state has the most discriminative effect on road type discrimination.

3.2 System application testing

After verifying the effectiveness of the algorithm model, to evaluate the comprehensive application performance of the designed vehicle navigation system integrating CART and DBSCAN in actual scenarios, a complete device-cloud collaboration system was experimentally deployed, and real vehicle testing was planned. The test aimed to evaluate the accuracy, real-time performance and reliability of the system in the real road environment, from data collection, cloud processing to navigation applications. The test fleet consisted of 5 vehicles equipped with different types of data collection terminals (3 via CAN routing and 2 via Android mobile app). A 1-week cycle test was conducted on a closed test route of about 50 km consisting of selected urban ring roads, expressways, and ordinary urban roads. The test route was manually surveyed in advance, and 12 curves (including 6 sharp curves and 6 gentle curves) and 8 typical road surface anomalies (such as sunken manhole covers, repaired joints, etc.) were marked as a verification benchmark for the system output results. The sharp turn is defined as the road section with a curve radius less than 100 m, and the slow turn is defined as the road section with a curve radius between 100 m and 300 m. All curve information is measured by high-precision differential GPS to measure the radius of curvature and record the start and end longitude and latitude. The pavement anomalies are divided into two categories according to the vertical acceleration peak value and the duration of vibration. The manhole cover depression type anomaly shows that the single vertical impact peak value is more than 1.5 m/s² and the duration is less than 0.5s, and the repair joint type anomaly shows that the continuous multiple vertical impact peak values are between 0.8 m/s² and 1.5 m/s² and the duration is 0.5s–1.5s. All abnormal points are recorded by manual patrol inspection.

During the test, the system received and processed more than 120 h of valid driving data. The experiment first collected the road type recognition results output by the cloud processing platform, compared them with the road type data provided by Gaode map, and calculated the recognition accuracy. The results are shown in Figure 11.

FIGURE 11

From Figure 11a, the correct identification mileage of the two road types under different time periods was basically proportional to the test mileage, but the correct identification mileage of ordinary urban roads during morning and evening peak hours decreased slightly. The accuracy data in Figure 11b showed that during the morning and evening peak hours with complex traffic and frequent starts and stops, the system’s recognition accuracy for ordinary urban roads was significantly lower than the early morning (98.1%) and midday off-peak hours (97.3%). For non ordinary urban roads, due to relatively simple road conditions and stable operating conditions, the recognition accuracy rate during each period remained above 98%, indicating strong robustness. The processing delay in Figure 11c was related to the system load and network conditions. Due to the increase in concurrent requests and network congestion during the morning and evening peak hours, the average delay rose to 235–250 m, but still met the real-time requirements (<500 m). During the early morning hours, it was as low as 192 m. In summary, the proposed system maintained high performance under different time periods and road conditions, especially stable performance on non-ordinary urban roads. Although the recognition accuracy of ordinary urban roads during peak hours fluctuated, it remained above 95% overall.

To explore the reasons for the decrease in accuracy of road recognition in ordinary cities during the morning rush hour, a case study was conducted on the misclassified samples. Randomly selecting 50 samples of ordinary urban roads that were mistakenly classified as “non ordinary urban roads” during the morning rush hour, analyzing their 28 dimensional feature distribution, it was found that 78% of the samples significantly deviated from the typical value range of ordinary urban roads in terms of the standard deviation of vehicle speed and lateral acceleration. Further tracing the driving states of the vehicles corresponding to these samples, it was found that they all occurred during the process of accelerating through the intersection after the red light started and stopped. The longitudinal acceleration, deceleration, and lateral steering coupling signals generated during this process have similarities with continuous curves at the feature level, leading to misjudgment of CART decision trees. To address this issue, the study adds two temporal correlation features on the basis of the original 28 dimensional features, namely, the cross-correlation coefficient between vehicle speed and lateral acceleration, and the short-term energy value of yaw rate, to distinguish between turning at intersections and continuous curves. The experiment shows that after adding the above features, the accuracy of ordinary urban road recognition during the morning rush hour increased from 95.8% to 97.1%, verifying the effectiveness of the optimization scheme. In addition, stratified analysis based on traffic flow density shows that the system’s recognition accuracy is 96.1%, 97.5%, and 98.2% in congested, slow-moving, and unobstructed states, respectively. The lowest accuracy is observed in congested states, further confirming the impact of complex start stop conditions on recognition performance.

In addition, the experiment also evaluated the system’s ability to detect precise information on the road surface. The system spatially matched the final accurate pavement information database generated after cloud clustering processing with 20 pre-marked reference points to calculate the detection rate and false alarm rate of the system. The results are shown in Figure 12.

FIGURE 12

From Figure 12, the test results showed that the system’s detection rate for sharp bends reached 100%, the detection rate for gentle bends was 91.7%, and the detection rate for road surface anomalies was 87.5%. The main missed detections occurred in individual gentle curves with small curvature and abnormal points on the road surface with unobvious vibration characteristics. At the same time, the system generated three false alarm points. After verification, they were all located near large intersections because the sensor signals of vehicle acceleration/deceleration and steering operations were similar to curve characteristics. After DBSCAN clustering processing, the data redundancy generated by repeated collection of multiple vehicles at the same location was effectively eliminated, and the average merging rate of data points reached 85%, which greatly improved the tidiness of the database. Finally, the experiment tested the navigation warning function from the perspective of end-user experience. Specifically, the test vehicle drove according to the navigation path provided by the system, and recorded whether the navigation app triggers a warning when approaching a curve or an abnormal point on the road in the database, the accuracy of the content of the warning prompt, and the advance amount of the warning. The warning advance was defined as the distance from the first voice/image prompt point to the center of the target point. The test results are shown in Table 3.

As can be seen, the test covers three types of warning scenarios: sharp turns, slow turns, and road surface abnormalities. The overall success rate of early warning is 89.2%, with an average early warning distance of 126 m. To address the issue of small sample sizes in each sub scenario, the Wilson method was used to calculate the 95% confidence interval. The success rate of the sharp turn warning in the scenarios of expressway and main road is 100% for both samples, with 95% confidence intervals of [34.2%, 100%] and [21.5%, 100%], respectively. In the main road scenario, the success rate of the slow bend warning was 100% with a 95% confidence interval of [29.2%, 100%] for 3 samples. In the ramp and auxiliary road scenario, the success rate was 60% with a 95% confidence interval of [17.3%, 92.7%] for 2.5 samples. The success rate of road surface anomaly warning was 100% with a 95% confidence interval of [29.2%, 100%] for three samples in the expressway scenario, and 75% with a 95% confidence interval of [25.1%, 96.0%] for four samples in the main road scenario. The success rate of slow bend warning in ramp and auxiliary road scenarios is relatively low, mainly due to the complex curvature changes and slope interference in this section of the road. Some curvatures are lower than the warning threshold set by the system, resulting in missed reports. The success rate of abnormal warning on the main road surface is 75%. The reason for this is that the foundation quality of the road surface in this section is good, and the vertical vibration characteristics of some abnormal points are not obvious. The difference from normal road surface paving is small, which makes the sensor signal amplitude ineffective in triggering the recognition conditions.

4 Conclusion

To accurately identify road types and aggregate accurate road surface information, a vehicle navigation system design method that integrates CART decision tree and DBSCAN algorithm was proposed. Experimental results showed that the accuracy, precision, recall and F1 score of the CART decision tree reached 96.8%, 97.2%, 97.3%, and 97.2%, respectively. The overall performance and training efficiency were better than the comparison model. Under the 30 m neighborhood radius of DBSCAN clustering, the average data point merging rate was 85.7%, and the noise elimination rate was controlled within 5%. In the system application test, the non-ordinary urban road recognition accuracy rate exceeded 98% in all periods, the overall detection rate of accurate road information was 92.5%, the navigation warning success rate was 89.2%, and the average warning advance distance was 126 m. Based on the above, the system achieved good results in recognition accuracy, processing efficiency, and early warning performance.

However, the recognition accuracy of ordinary urban roads in peak hours fluctuates slightly under the influence of complex start-up and stop conditions, and some slow bends and road surface anomalies are missed. The follow-up research will focus on multi-source sensor data such as visual images and millimeter wave radar. The convolutional neural network is used to automatically extract spatial-temporal features to reduce the dependence on manual statistical features. The domain adaptive technology is used to improve the generalization ability of the model between different vehicle types and different urban road conditions, and further optimize the recognition robustness and deployment flexibility of the system under complex conditions.

In addition, the sample size of the early warning test for the central molecular scenario in the system application testing is relatively small. Although the Wilson confidence interval is used to quantify the uncertainty of the proportional estimation, the statistical reliability of the conclusion still needs to be further verified after expanding the testing scale. In the future, more real vehicle tests will be conducted to accumulate early warning data in various scenarios.

References

• 1

  BajalE.KataraV.BhatiaM.HoodaM. (2022). A review of clustering algorithms: comparison of DBSCAN and K-mean with oversampling and t-SNE. Recent Pat. Eng.16 (2), 17–31. 10.2174/1872212115666210208222231

  • CrossRef
  • Google Scholar

• 2

  ChenK. (2022). Road roughness recognition based on lidar. J. Phys. Conf. Ser. IOP Publ.2278 (1), 012008. 10.1088/1742-6596/2278/1/012008

  • CrossRef
  • Google Scholar

• 3

  ChenG. (2025). Intelligent recognition of financial fraud based on CART decision tree. Int. J. Inf. Commun. Technol.26 (11), 1–20. 10.1504/IJICT.2025.146100

  • CrossRef
  • Google Scholar

• 4

  ChenL.HeX.ZhaoX.LiH.HuangY.ZhouB.et al (2022). Gocomfort: comfortable navigation for autonomous vehicles leveraging high-precision road damage crowdsensing. IEEE Trans. Mob. Comput.22 (11), 6477–6494. 10.1109/TMC.2022.3198089

  • CrossRef
  • Google Scholar

• 5

  DaiS.LiP. (2024). Calculation model of distributed photovoltaic carrying capacity for 110 kV power supply area based on CART decision tree. Distrib. Energy9 (3), 82–88. 10.16513/j.2096-2185.DE.2409310

  • CrossRef
  • Google Scholar

• 6

  LeiT.SellersT.LuoC.CaoL.BiZ. (2024). Digital twin‐based multi‐objective autonomous vehicle navigation approach as applied in infrastructure construction. IET Cyber‐Systems Robotics6 (2), e12110. 10.1049/csy2.12110

  • CrossRef
  • Google Scholar

• 7

  LianZ.ZengQ.WangW.XuD.MengW.SuC. (2023). Traffic sign recognition using optimized federated learning in internet of vehicles. IEEE Internet Things Journal11 (4), 6722–6729. 10.1109/JIOT.2023.3312348

  • CrossRef
  • Google Scholar

• 8

  LiuJ.LiuH. O.ChenH. Y.MaoF. H. (2023a). Road types identification method of unmanned tracked vehicles based on fusion features. Acta Armamentarii44 (5), 1267–1276. 10.12382/bgxb.2022.0038

  • CrossRef
  • Google Scholar

• 9

  LiuW.LiuY.BucknallR. (2023b). Filtering based multi-sensor data fusion algorithm for a reliable unmanned surface vehicle navigation. J. Mar. Eng. and Technol.22 (2), 67–83. 10.1080/20464177.2022.2031558

  • CrossRef
  • Google Scholar

• 10

  MoY.ZhangP.ChenZ.RanB. (2022). A method of vehicle-infrastructure cooperative perception based vehicle state information fusion using improved kalman filter. Multimedia Tools Applications81 (4), 4603–4620. 10.1007/s11042-020-10488-2

  • CrossRef
  • Google Scholar

• 11

  NandhuC.GopiA.NithamdharB.LaxmiprasannaC.SrihariD.JitendraA.et al (2026). IoT-based vehicle tracking with accident alert system. Int. J. Res. Publ. Eng. Technol. Manag. (IJRPETM)9 (2), 486–494. 10.15662/IJRPETM.2026.0902001

  • CrossRef
  • Google Scholar

• 12

  NiuX.PengY.DaiY.ChenQ.GuoC.ZhangQ. (2022). Camera-based lane-aided multi-information integration for land vehicle navigation. IEEE/ASME Trans. Mechatronics28 (1), 152–163. 10.1109/TMECH.2022.3192985

  • CrossRef
  • Google Scholar

• 13

  PangY.ZhuX.HeT.LiuS.ZhangZ.LvQ.et al (2024). AI‐assisted self‐powered vehicle‐road integrated electronics for intelligent transportation collaborative perception. Adv. Mater.36 (36), 2404763. 10.1002/adma.202404763

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 14

  RafikaC. S.ReynandaR. M.SariA. P. (2025). Penerapan decision tree CART untuk klasifikasi risiko gagal studi mahasiswa. J. Mhs. Tek. Inform.4 (2), 37–46. 10.35473/jamastika.v4i2.4145

  • CrossRef
  • Google Scholar

• 15

  TangC.WangC.ZhangL.ZhangY.SongH. (2024). Vehicle heterogeneous multi-source information fusion positioning method. IEEE Trans. Veh. Technol.73 (9), 12597–12613. 10.1109/TVT.2024.3393720

  • CrossRef
  • Google Scholar

• 16

  WangY.SunR.ChengQ.OchiengW. Y. (2023). Measurement quality control aided multisensor system for improved vehicle navigation in urban areas. IEEE Trans. Industrial Electron.71 (6), 6407–6417. 10.1109/TIE.2023.3288188

  • CrossRef
  • Google Scholar

• 17

  WeiK.YuL.XuF. (2024). Integrated Road Information Perception Framework for Road Type Recognition and Adaptive Evenness Assessment[C]//WCX SAE World Congress Experience 288443. 10.4271/2024-01-2041

  • CrossRef
  • Google Scholar

• 18

  YangY.QianC.LiH.GaoY.WuJ.LiuC. J.et al (2022). An efficient DBSCAN optimized by arithmetic optimization algorithm with opposition-based learning. Journal Supercomputing78 (18), 19566–19604. 10.1007/s11227-022-04634-w

  • CrossRef
  • Google Scholar

• 19

  YangH.WuJ.HuZ.LvC. (2023). Real-time driver cognitive workload recognition: attention-Enabled learning with multimodal information fusion. IEEE Trans. Industrial Electron.71 (5), 4999–5009. 10.1109/TIE.2023.3288182

  • CrossRef
  • Google Scholar

• 20

  YangB.ZhuX.PengC.ZhouL.WangF.LiuZ.et al (2025). Synchronized measurement of electromechanical responses of fabric strain sensors under large deformation. Smart Wearable Technol.1, A7. 10.47852/bonviewSWT52026022

  • CrossRef
  • Google Scholar

• 21

  ZhangH.WangZ.XiaW.NiY.ZhaoH. (2022). Weighted adaptive KNN algorithm with historical information fusion for fingerprint positioning. IEEE Wireless Communications Letters11 (5), 1002–1006. 10.1109/LWC.2022.3152610

  • CrossRef
  • Google Scholar

• 22

  ZhaoH.LiuJ.ChenX.CaoH.WangC.LiJ.et al (2024). Information monitoring and adaptive information fusion of multisource fusion navigation systems in complex environments. IEEE Internet Things J.11 (14), 25047–25056. 10.1109/JIOT.2024.3391872

  • CrossRef
  • Google Scholar