Application of Multi-Target Visual Tracking Method Combined with HRNet in Complex Dynamic Scenes

In crowded streets or markets, the tracking task becomes extremely challenging due to mutual occlusion, similar appearance, and rapid motion changes between targets. The paper first processes the input video frame, then uses the same human frame detector as in Simple Baseline, propagates the detected frame to adjacent frames by propagating the predicted key points according to the optical flow calculated by FlowNet 2.0, and performs non-maximum suppression. Then, pose estimation is performed within the detected human frame, and key point detection is performed using HRNet. Correspondences between keypoints in adjacent frames are computed using a greedy matching algorithm based on the similarity between the keypoints within a frame and the keypoints propagated from adjacent frames according to optical flow. Finally, the feature extraction capability of HRNet and the optimized re-identification branch are combined to achieve multi-target tracking in complex dynamic scenes. The research results show that the average tracking success rate of the multi-target visual tracking method combined with HRNet reaches 97.555%, and the number of loss times in multiple test scenarios is 0. Compared with the traditional method based on mean shift, the average IDF1 value and the highest MOTP value of this method reach 0.9115 and close to the full score respectively. Therefore, this paper provides an effective solution to the problem of multi-target tracking in practical scenarios.