Real-Time Detection and Tracking Framework Using Extended Kalman Filter BoT-SORT in Uncertainty Mixed-Traffic

Abstract

Perception systems play a crucial role in real-time decision-making in intelligent transportation, particularly in uncertain traffic. Challenges such as dynamic movement, unpredictability, occlusion, and ambiguous interactions necessitate the development of adaptive detection and tracking frameworks. To address these issues, we present the uncertainty mixed-traffic (UMT-Dataset), an extension of the MXT-Dataset, tailored to address dynamic object behavior in mixed-traffic environments. We also propose the YOLOv10-UMT framework, which integrates YOLOv10n with a modified bag-of-tricks for re-identification + simple online and real-time tracking (BoT-SORT) algorithm enhanced by an extended Kalman filter (EKF) and a noise scaling adaptive (NSA) mechanism. This method enhances BoT-SORT's ability to estimate object positions more reliably under uncertain conditions. The EKF integration can handle nonlinear trajectories more accurately, whereas the NSA can adaptively adjust measurements for detection. Experimental results show that integrating YOLOv10n with modified BoT-SORT using EKF+NSA significantly improves the precision and efficiency. This method achieves HOTA 42.064, MOTA 22.868, and IDF1 46.324 with an inference time of 2668 ± 37.01 ms. Evaluations on datasets of varying sizes (1600, 2000, and 2500 images) further confirm the robustness of EKF+NSA, supported by 95% confidence intervals (CI), inference time standard deviations, and computational cost analysis. Additionally, YOLOv10n trained on the UMT-Dataset outperformed YOLOv9t and YOLOv11n, achieving mAP@0.5 of 0.858, precision 0.868, recall 0.781, F1-score 0.82, and speed of 555.56 FPS. The proposed method is effective for adaptive detection and tracking in uncertain traffic, prioritizing accuracy, time efficiency, and contributing to a reliable perception module in real-world intelligent transportation systems.