GBR-Sort: A Robust and Efficient Online Multi-Object Tracking Method for Campus Safety

Deep learning-based object detection has driven the development of multi-object tracking (MOT). However, the existing MOT methods are mainly aimed at regular motion, which makes it difficult to deal with the common issues of irregular motion and occlusion in campus safety. To deal with these issues, we propose a novel robust and efficient online algorithm named GBR-Sort. Firstly, a lightweight OSNet model is used to extract appearance features, which are fused with extended BIoU cosine-distance using a new strategy. Secondly, to improve the accuracy of motion estimation and fix the error accumulation of filter parameters, we design a motion model by combining the linear motion estimation with the Kalman filter. Finally, we introduce a multi-level nested cascade matching method, which effectively reduces the probability of mismatching by simultaneously considering the detection confidence and the size of expanded BIoU during the matching process. In addition, to enrich the MOT dataset in campus safety, we build a campus monitoring MOT dataset named School MOT Data (SMD), covering various scenes, motion states and illumination conditions. The experimental results on SMD show the superiority of the proposed method, with HOTA, MOTA and IDF1 reaching 75.823%, 90.52% and 87.622%, respectively, even in challenging conditions like the soccer field. Our method achieves a good trade-off between accuracy and efficiency, making it suitable for online MOT tasks in real campus scenes.