Adaptive vehicle counting method using YOLOv10 integrated with BoostTrack at symmetrical traffic light intersections
Abstrak
ABSTRACT The increasing volume of vehicles in urban areas creates complex traffic patterns, particularly at symmetrical intersections with mixed-traffic characteristics common in Indonesia. Accurate vehicle counting is essential for adaptive traffic signal systems, as real-time traffic density estimation can reduce waiting times, lower fuel consumption, and improve intersection efficiency. However, low illumination, inter-vehicle occlusion, and high traffic density remain major challenges for existing vehicle counting approaches. This study aims to develop a robust and efficient vehicle counting system for symmetrical intersections under mixed-traffic conditions. We introduce the Crossjunction Traffic Dataset (CJT-D), a novel annotated benchmark consisting of 5,000 images with 111,531 object annotations collected under four illumination conditions: morning, noon, afternoon, and night. Furthermore, we propose a unified framework integrating YOLOv10n as the primary detector with the confidence-aware BoostTrack tracker and an adaptive zone-based counting mechanism tailored to symmetrical intersection geometry. The detection model trained on CJT-D achieves a mean average precision of 0.952, an F1-score of 0.91, and an inference speed of 1.7 milliseconds per image. The tracking module achieves a higher-order tracking accuracy of 56.82, multiple objects tracking accuracy of 58.49, and identity consistency score of 72.70. Vehicle counting accuracy reaches 100% in the morning, 84–91% in the afternoon and evening, and 66–73% at night. These results confirm that the proposed framework delivers accurate, efficient, and occlusion-tolerant real-time vehicle counting at symmetrical intersections, offering strong potential for a computer vision-based adaptive traffic light control system.
Identifier
- DOI: https://doi.org/10.1016/j.fraope.2026.100720
- OpenAlex: https://openalex.org/W7169827627