A Cascading Driven Multi-Object Tracking Algorithm Based on StrongSORT

Multi-object tracking(MOT) algorithms aim to accurately locate and track multiple objects across consecutive video frames while maintaining consistent object identities. Currently, mainstream MOT algorithms often leverage appearance information to assist filtering algorithms in accomplishing MOT tasks. However, this idea primarily relies on appearance information of the objects. When objects repeatedly enter and exit the lens or undergo appearance changes, the objects may have ID switches, leading to the failures of the MOT task. Based on this, this paper proposes a cascading driven multi-object tracking algorithm based on StrongSORT. Firstly, a lower part of body appearance information matching module and a skeleton information matching module are designed to address the issue of object appearance changes. These two modules are independent and are integrated with the original matching module in a cascading method through computing the joint costs respectively. Secondly, this paper replaces the detector with YOLOv7 and employs OSnet to extract object appearance features. The best features are preserved for a longer time by a long-term memory mechanism proposed in this paper, enhancing the robustness of the tracker when the object repeatedly enters and exits the lens. Additionally, a custom dataset is built to test the algorithm proposed in this paper. Compared to the original algorithm, the number of ID switches has been reduced by 20%, the identity recognition performance has improved by 7.07%, and the tracking performance has increased by 6.11%. In conclusion, this method successfully addresses the issue of MOT task failure caused by the tracked object repeatedly entering and exiting the lens or undergoing appearance changes.