Person Re-identification Based On Spatio-temporal Information Fusion

With the increasing demand in the field of video surveillance,pedestrian recognition technology has received extensive attention.Pedestrian re-identification aims to use computer vision technology to determine whether there is a specific pedestrian in an image or video sequence,that is,to give a specific monitoring pedestrian image and match the image of the pedestrian under the camera.This technology is designed to compensate for the current fixed camera vision.Limitations.In the actual video surveillance environment,lighting changes,camera angle changes,complex background and severe occlusion factors cause the pedestrian recognition task to face enormous challenges.The traditional single-frame image-based method is difficult to adapt to the actual scene,and the video can more feature information is provided from the perspective of time and space,which is helpful for pedestrian matching and re-identification.Therefore,the research on pedestrian recognition based on video sequences has emerged.This paper analyzes the various methods in the relevant fields at home and abroad for the problem of pedestrian recognition,and summarizes its advantages and disadvantages.The main research idea of the subject is: input video sequence,the spatial feature of image level is obtained by 2D convolutional neural network,and the method of time series modeling is used to effectively combine the spatial and temporal features to obtain the feature information of the video level.Finally,the feature distance is measured to re-identification.The main research content includes the following parts:(1)Target detection network extracts video key frames.Considering that many redundant information in the video sequence will interfere with the matching accuracy of pedestrian recognition,such as pedestrian occlusion and low resolution.Based on the spatio-temporal information fusion network,the target detection network can be effectively filtered out,and the image frames with higher confidence can be screened out,and discriminative image frames can be selected to reduce the interference caused by pedestrian occlusion and other factors,thus improving pedestrian recognition.Matchingperformance.(2)Convergence of spatial pyramids and temporal attention mechanisms.On the basis of the fusion target detection network and the spatial depth feature,in the time series modeling part,the spatial pyramid idea is incorporated,and spatial multi-scale attention is paid to each frame image,and the time domain convolution fusion time-series context information is combined.The mechanism solves the defect of different input pictures,and the multi-scale feature extraction achieves better image recognition accuracy.In addition to the attention mechanism,this paper also studies Temporal pooling and LSTM time series modeling methods.After analysis and comparison,the timing mechanism of attention mechanism is better than the other two.(3)Classification and distance measurement loss function fusion.The loss function is one of the key steps in the pedestrian recognition task.This paper considers the category label information and the feature distance information separately,and combines the two loss functions to achieve better re-recognition performance.(4)Pedestrian re-identification network based on 3D convolution network.The above-mentioned spatio-temporal feature information fusion method first extracts the spatial features of the image level,and then fuses according to the time series features,and the 3D convolution network directly processes the three-dimensional data of the video sequence,and can simultaneously extract the spatial and temporal feature information,and the method is simple,effective.