Research on Video Pedestrian Re-ID Method Based on Global Branching and Channel Division

Pedestrian re-identification is a new technology that has emerged in the field of intelligent detection and analysis in recent years, and is increasingly attracting the attention of researchers. Compared with image-based pedestrian. Re-identification, Video-based pedestrian re-identification has more information. However it is faced with more gradient disappearance or gradient explosion, which makes low efficiency, poor generalization ability, thus leading to low model efficiency. Therefore, this paper proposes a video pedestrian re-identification method based on Channel Division and Global Branching. Firstly, the backbone network of ResNet50 is introduced, and the global feature and Channel Division are combined to comprehensively take into account the global information and feature extraction deepening. Then, the SEWeight module is introduced, whose compression and stimulation parts can be used to encode global information and adaptively recalibrate the channel relationship. Multiple loss functions and activation functions are also used to optimize the model. Finally, extensive experiments have verified the effectiveness of our method, and our method has achieved the most advanced results on two major datasets: Mars and iLDS-VID.