Pedestrian multi-object tracking algorithm based on improved YOLOX and multi-level data association

Object tracking is a fundamental problem in the field of computer vision. Pedestrian multi-object tracking has broad application prospects in many fields such as intelligent surveillance and intelligent transportation. However, frequent occlusion and scale change exist in the actual tracking scene, which brings great challenges to the multi-object tracking algorithm. In order to further improve the tracking accuracy, on the basis of DeepSORT, a pedestrian multi-object tracking algorithm based on improved YOLOX and multi-level data association was proposed. For the detector, in order to enhance the feature expression ability of the network and improve the detection accuracy, the ECA channel attention module and the ASFF adaptive feature fusion module were introduced into the YOLOX skeleton network and the neck network respectively. For identification features, in order to reduce the number of false matches in the data association step and improve tracking efficiency, the lightweight OSNet re-identification network and NSA Kalman filter were used to obtain target features. For data association, in order to reduce the number of identity switching and avoid target loss, the detection and tracking were classified. The different similarity calculation methods were used to realize multi-level data association based on detection confidence and trajectory state. The experimental results show that compared with the algorithm that simply combines YOLOX and DeepSORT before improvement,the introduction of ECA module and ASFF module in YOLOX can reduce the number of false detections. The reduction can be up to 17% when using YOLOX-s model. The feature extraction method combining OSNet model and NSA Kalman filter can improve the tracking stability, the IDF1 index is increased by 0.77%, and the IDSW is reduced by 947. The multi-level data association algorithm based on detection confidence and trajectory state can significantly improve the tracking performance, and the MOTA index is increased by 3.36%. The results of MOTA on MOT17 and MOT20 test sets are 80.4% and 77.7%, and IDF1 are 78.4% and 76.7%. Compared with other advanced algorithms, the proposed pedestrian multi-object tracking method achieves a better balance between tracking accuracy and tracking speed, which can provide reference for online pedestrian multi-object tracking applications in industry. © 2024, Central South University Press. All rights reserved.