Oblique-View Video Tracking and Density-Based Counting: Accurate Counting of Late-Stage Rapeseed Seedlings for Breeding Assessment

Bowen Luo¹Yuang Yang¹Kuanyan Zhang¹Xuan Lv¹Yujie Liu¹Yicheng Yang¹Fugui Zhang¹Lu Liu¹Gang Zhang²Xiaole Wang¹Zhenchao Wu^1*

• ¹Anhui Agricultural University, Hefei, China
• ²Anhui Science and Technology University, Bengbu, China

Accurate counting of rapeseed seedlings at the late seedling stage is critical for yield estimation and field management, yet traditional manual counting is inefficient and labor-intensive. To address this, this study proposed a novel video tracking and counting method (CropTriangulator) for late-stage rapeseed seedlings based on oblique view and target density distribution, utilizing smartphone-captured videos to achieve row-based accurate counting. The method integrated three core components: target detection, adaptive clustering, and multi-object tracking. First, YOLOv11 models of different scales were trained and compared, with the nano-scale model selected for its balanced performance in detection accuracy (AP0.5:0.95 of 0.811 for 45° view) and inference speed (4.6 ms per frame). Second, an adaptive DBSCAN (AdapDBSCAN) algorithm was proposed to eliminate rapeseed seedlings in non-target regions by dynamically adjusting clustering parameters (eps and min_samples) based on local target density, addressing perspective-induced distortion (smaller in distance and larger in foreground). Finally, the SORT algorithm was employed to track and count the extracted target seedlings, with ID uniqueness ensured by permanent marking of IDs when seedlings cross the video frame boundary. Experimental results on 20 test videos (10 for 45° oblique view and 10 for 90° vertical view) showed that the CropTriangulator method achieved an average counting accuracy of 97.13% for the 45° view, which was 14% higher than that for the 90° view (82.94%). The R-squared (R2) of row-based counts for the 45° view reached 0.917, indicating strong stability. Compared to fixed-parameter DBSCAN, AdapDBSCAN reduced over-counting by filtering non-target seedlings, with the SORT algorithm exhibiting superior tracking performance (ID switch rate of 8.47%) compared to DeepSORT (36.05%). This study demonstrates that the 45° oblique view was optimal for rapeseed seedling counting, and the proposed CropTriangulator method provides a low-cost, efficient solution for automated row-based counting in complex field environments, supporting precise yield estimation and orchard management decisions. The video comparing the effects of the CropTriangulator method is available at: https://github.com/Possibility007/Comparison-of-counting-results.git