A Pedestrian Re-identification Algorithm Based on 3D Convolution and Non_Local Block

Abstract

In the application of video-based pedestrian re-identification, introduced deep learning method to learn feature representation of pedestrian. In order to improve feature quality, introduced 3D convolution block as backbone network to aggregate temporal and spatial features; for issue of human body occlusion in video frames, introduced Non_Local block to capture long distance dependence between frames, and eventually eliminate the impact of occlusion. Optimal embedding scheme of 3D convolution and Non_Local block in backbone network is designed via experiments, and has proved that rich features of pedestrian can be extracted from video frames by this solution, which helps to improve the accuracy of re-identification.