People tracking accuracy improvement in video by matching relevant trackers and YOLO family detectors

Abstract

The tracking-by-detection paradigm is widely used for people multi-object tracking tasks. Up to now, there exist many detectors and trackers, many evaluation benchmarks, which neces sitates the use of relatively uniform estimation methods and metrics. It leads to necessity to choose better combined models of detectors and trackers. To solve this task, we developed a comprehensive performance evaluation methodology for estimation of people tracking accuracy and real-time by using different detectors and trackers. We conducted experiments by choosing the official pre-trained models of YOLOv5, YOLOv6, YOLOv7, YOLOv8 with representative BoTSORT, ByteTrack, DeepOCSORT, OCSORT, StrongSORT trackers under two benchmarks of MOT17 and MOT20. Detailed metrics in terms of error and speed such as higher order track ing accuracy and frames per second were analyzed for the combinations of detectors and track ers. It is concluded that the OCSORT+YOLOv6l model has the best comprehensive perfor mance and the combination of OCSORT and YOLOv7 has the best average performance under MOT17 and MOT20