An online multi-camera multi-vehicle tracking using lightweight YOLO11 and improved association strategies

Multi-camera multi-vehicle tracking (MCMVT) is essential for intelligent transportation and urban security but is hindered by similar vehicle appearances, varying viewpoints, and occlusions. While existing offline algorithms offer high accuracy, their heavy computational load and lack of real-time performance limit practical use. To overcome these issues, this paper proposes an efficient online MCMVT method that improves tracking accuracy in complex environments and is suitable for edge devices. The method utilizes a lightweight YOLO11s detector to reduce detection time without compromising accuracy. The single-camera multi-vehicle tracking (SCMVT) algorithm is enhanced with a joint matching strategy based on cosine feature distances and IoU, combined with Exponential Moving Average (EMA)-based dynamic updates and small-target rejection, substantially boosting tracking accuracy and robustness. For cross-camera trajectory association, a hierarchical clustering algorithm based on cosine distance and Dunn’s index automatically optimizes clustering parameters to enhance matching precision. Additionally, leveraging SCMVT results maintains trajectory ID stability, reducing ID switching by 52% and lowering computational overhead. Experimental results demonstrate that the proposed method achieves an IDF1 score of 81.64 on the S02 scenario of the CityFlowV2 dataset with an average tracking time of 8 ms per frame and delivers real-time performance at 3 FPS on the NVIDIA Jetson AGX Orin edge device, confirming its efficiency and practicality.