Tracking Algorithm Based on Video Person Reidentification and Spatiotemporal Feature Fusion

Multiobject tracking algorithms are frequently affected by the problem of the exchange of pedestrian identity in real congestion situations. To solve this problem, this study proposes a joint network that integrates target detection and person reidentification. Additionally, a track scoring mechanism is introduced to integrate the reidentified feature and time information. By collecting candidates from the detection results and tracking prediction results, the tracking prediction information and reidentified feature information of pedestrian targets can complement each other. To solve the problem of detecting small targets in video images, this study improves the ResNet-34 network by combining the deep aggregation network on the backbone network and replacing the traditional residual block with a multiscale convolutional network to focus on small targets and improve the detection accuracy. In this study, experiments were conducted on the multiobject tracking datasets MOT16, MOT17, and MOT20. The corresponding multiple object tracking accuracy (MOTA) of the proposed network reaches 74.7, 73.7, and 66.4, respectively, and the conversion durations of pedestrian identity are 210, 209, and 1403, respectively. The results reveal that the proposed network has good detection and tracking performances.