Research On Video Person Re-identification Algorithm Based On Convolutional Neural Network

With the national promotion of "Safe city","Skynet project" and "Snow bright project" to build multi-level video monitoring and networking applications,the number of monitoring video is increasing day by day.In terms of security,it will be very difficult to track and monitor people in front of a large number of videos by relying on human resources alone.Person ReIdentification(ReID)is the key technology to solve this problem.It is a very important and challenging topic in the field of computer vision.Because of the complex environment in the actual application scene,the camera angle of view and light and other conditions change greatly,and it is easy to have multiple people occlusion phenomenon.Therefore,how to extract the comprehensive and discriminative person features is the key problem of ReID task.With the development of convolution neural network,using convolution neural network model to extract person features can effectively improve the accuracy of ReID.In this study,the model and algorithm of ReID based on convolution neural network are studied.The main research work is as follows:(1)In view of the current dataset image misalignment,the existing ReID model is difficult to obtain a more comprehensive person descriptor.This study proposes ReID model based on bilinear CNN and part features combined with the fine-grained recognition framework.In this model,the bilinear framework is composed of a simplified residual network,and the feature is further segmented horizontally by using the branch network to extract more fine-grained features,so that the fused person has both full-body features and fine-grained features.In order to reduce the distance between the same person and expand the distance between different person,the network model is optimized by using the cross entropy loss after label smoothing and the triplet loss of hard to separate negative samples.The validity of the algorithm is verified on the datasets Market-1501 and CUHK03.The results show that the proposed algorithm is better than the traditional algorithm and other similar ReID algorithms in performance.(2)In view of the shortcomings of global local descriptor alignment network in feature extraction and different part weighting,this study proposes a ReID model based on attention mechanism and adaptive weighting.This method uses the human key point detection network to extract the person area,divides the person into head,upper body and lower body,and then uses the feature extraction network embedded in the attention mechanism to extract the whole body and three parts.The attention mechanism can select the features by weighting the useful channel features to improve the recognition ability of the extracted features of the network.An adaptive weighting module is designed to weigh the three parts of person more reasonably,and further improve the performance of the model.The performance comparison and validation of the algorithm are carried out on the datasets DukeMTMC-reID and Market-1501.The experimental results show that the embedded attention network can extract features with higher discrimination,and the adaptive weighting module can also improve the performance.