Research On Video Person Re-identification Based On Deep Learning

In recent years,with the great success of deep learning in the field of computer vision,intelligent monitoring systems based on deep learning have attracted widespread attention from researchers.Pedestrian re-identification,as an important part of the intelligent monitoring system,has a huge role in finding lost people,investigating criminal cases,and urban safety management,and has become a new research hotspot.The main task of video pedestrian re-identification is to determine whether the pedestrians captured by different cameras are the same identity based on the video sequence.However,due to different camera capture angles of pedestrians,different lighting conditions,and occlusion between objects or pedestrians,how to solve the problem of accurate pedestrian re-identification is still a huge challenge.This thesis summarizes and researches current video person re-identification methods and makes improvements to existing problems.Most current video pedestrian re-recognition methods need to manually extract optical flow maps in advance to calculate pedestrian motion features,while manually extracting optical flow features is time-consuming and laborious and optical flow maps occupy a lot of disk space,which is not conducive to practical applications.To solve this problem,a video pedestrian re-identification method based on optical flow guided features is proposed in this paper.First convolutional neural networks are used to extract the spatial appearance features of pedestrian images,and optical flow guided features are calculated to describe pedestrian motion information based on the depth feature map.Then pedestrian spatial appearance features and temporal motion features are combined to obtain video-level pedestrian feature description vectors.Finally,the Euclidean distance between the feature description vectors is calculated to determine whether the pedestrians in the two image sequences have the same identity.Experiments show that thismethod can effectively simplify the network structure and achieve higher recognition accuracy.At present,most methods obtain as much information as possible to achieve pedestrian re-recognition.However,not all information can make enough contributions to pedestrian re-recognition.The attention mechanism can extract useful pedestrian information to suppress useless information,and apply more computing resource to key information to improve the accuracy of pedestrian re-identification.Based on the attention mechanism,we propose a video person re-identification method based on spatio-temporal attention.Firstly,the residual attention network is used to extract the effective information in the channel and spatial dimensions of the feature map,and the temporal attention model is used to realize the temporal dimensional effective feature extraction.Then effective information in the three dimensions of channel,space,and time are synthesized to obtain a more discriminative and more comprehensive video-level pedestrian feature description vector.Finally,the loss of multi-task is used to train the network.Experiments on datasets i LIDS-VID,PRID-2011 show that the accuracy rate is higher than most current methods,which proves the effectiveness of the method.