Person Re-identification Method Based On Video Attention

As the future application of person re-identification is very extensive,it has been one of the research hotspots in the field of computer vision in recent years.With the development of deep learning and the improvement of computer equipment performance,more and more researchers have shifted their focus from image-based to video-based person re-identification.However,except the traditional problem like the changing of lighting,occlusion,viewing angle and background,there is another important problem in this field,that is,the amounts of calculations are doubled because of video sequences.In this article,how to effectively improve the accuracy of the algorithm while reducing the amount of calculation is the main issue.In order to achieve the goals,we first propose a lightweight 3D global context attention network structure,which is embedded in the various stages of the residual network at multiple levels.The lightweight attention structure would not bring much additional calculations when embedding the convolutional network,and the multilevel embedding residual network can obtain multi-level spatio-temporal information and better extract video features.Subsequently,this paper proposes a random erasure method for the video-based person re-identification,which can improve the algorithm’s robustness to occluded videos without incurring additional calculations.In addition,this paper also introduces a gradient centralization method in the optimizer,which effectively improves the stability and accuracy of network training.This paper verifies the proposed algorithm on the public large-scale video person re-identification data set Mars.Compared with the strong baseline,our method effectively improves the accuracy while reducing the GPU memory usage by half during training,with Rank-1 reaching 90.8% and m AP reaching 85.2%.Experiments shows that the algorithm can effectively improve the accuracy of the algorithm while greatly reducing the hardware requirements for training.