Spatiotemporal Attention On Sliced Parts For Video-based Person Re-identification

Person re-identification refers to identify persons from different monitoring views by extracting persons’visual features.As a key technology of computer vision,person re-identification has a great development prospect.However,there are many challenges for person reidentification in the practical application.Due to the occlusions of persons’body,background changes and so on,this technology faces complex images.What’s more,the images captured by monitor equipment are usually continuous,it’s hard to choose specific image for person reidentification because it will bring tremendous manpower cost.To solve the above problems,this thesis uses continuous person image sequences intercepted from surveillance video for video-based person reidentification and proposes a new video-based person re-identification algorithm.The main work of this thesis includes the following three points:1.In this thesis,a feature map slice method based on key points clustering of human skeleton is proposed to obtain local features of person images.Compared with extracting the global features of the person images directly,this method can retain the detail information of the person images and represent the person images more accurately.2.In order to further optimize the spatial feature extraction of person images,this thesis proposes a spatial attention mechanism based on sliced parts.By generating spatial attention matrix,the effective regions of person can be found in the image,so as to obtain more representative spatial feature representations of person images.3.In order to solve the problem that the quality of images differs greatly in the image sequences,this thesis proposes a temporal attention mechanism based on sliced parts,which gives temporal weight to each image region in the image sequences and optimizes the feature fusion of person images.In addition,this thesis also combines multiple types of loss functions to jointly supervise the training process of the overall model.This model achieves excellent performance in the popular video-based person reidentification datasets including PRID2011,ILIDS-VID,and MARS.