Limited Resources Multi-Object Tracking by Utilizing Motion Cue from the Meanshift Algorithm

Multi-object tracking has various applications in intelligent surveillance systems. Substantial advancements have been made to enhance the accuracy of multi-object tracking. Nevertheless, many existing approaches struggle with real-time tracking in resource-constrained environments such as surveillance centers. While it's possible to get close to real-time tracking with low resources utilizing motion-based trackers, their framerates are frequently limited by detectors. In this article, we introduce a technique that reduces the frequency of object detectors and association algorithms calls in SORT-like trackers by performing target detection and data association every several frames, predicting the positions of bounding boxes between target detections using motion cues derived from the meanshift algorithm. By this mean, we achieved real-time multi-object tracking with an average end-to-end framerate of 38.09 fps on the MOT17 dataset on a single mobile RTX4060 65W, which is 230% of the original SORT implementation while maintaining most of the precision metrics evaluated in IDF1 and MOTA, significantly increasing the tracking framerate in resource-constrained scenarios.