Research On Video Person Re-identification Method Based On Spatial-temporal Attention Mechanism

As one of the latest studies in computer vision,Person Re-identification plays an extremely important role in public security,such as security monitoring and criminal investigation.With the development of technology and the advancement of deep learning,Person Re-identification has also received widespread attention from all walks of life.Although Person Re-identification methods based on deep learning has greatly improved over some traditional methods,there are still many difficulties in Person Re-identification.In reality,the environment where the pedestrian is in is very complicated.Therefore,when extracting the feature from the input video sequence,effectively filtering background noise and extracting more discriminative pedestrian feature from pedestrian sequences with complex backgrounds puts forward high requirements for the feature extraction network.In addition,the problem that pedestrian may be occluded in the pedestrian sequence also poses a great challenge to Person Re-identification.In view of the above difficulties,this paper has completed the following researches: How to effectively eliminate the impact of complex backgrounds in Person Re-identification;How to improve the network's feature extraction capability for input video sequence;How to solve the problem of pedestrians being obscured;The overall optimization of the network.In this paper,two auxiliary sequences are used to guide the original pedestrian video sequence to solve the problem caused by complex background in original sequence.The optical flow sequence and the mask sequence are obtained from the original sequence through the optical flow algorithm and the semantic segmentation algorithm.Both the optical flow sequence and the mask sequence can distinguish between pedestrian foreground and background noise very well,so they are both used as the auxiliary sequences of the original input sequence.The original sequence and auxiliary sequences are used as the input of the Person Re-identification network together to make the auxiliary sequences guide the feature extraction network to get more effective features from the pedestrian video sequence.In terms of network structure,in order to extract more discriminative pedestrian feature from the input pedestrian video sequence,this paper will compare several recent feature extraction network structures and combine their advantages to design a new feature extraction network that can perform more effective feature extraction on input pedestrian video sequences.Furthermore,this paper also embeds the spatial attention mechanism and the temporal attention mechanism into the Person Re-identification task and studies the video Person Re-identification method based on the spatial-temporal attention mechanism to solve the problem of pedestrians being occluded.By embedding the spatial-temporal attention mechanism,the neural network can effectively amplify the part of the pedestrian sequence that is beneficial to pedestrian reidentification and weaken the part that is not useful or even negative for the recognition when performing Person Re-identification tasks.Let the neural network has the visual attention mechanism like a human,know “where to look” and “when to look”.And this paper will prove the effectiveness of the Person Re-identification network based on the spatial-temporal attention mechanism through a large number of experiments.In terms of network training and testing,this paper will use a more effective loss function for Person Re-identification through experimental comparison to make the network better trained.In addition,the similarity aggregation strategy is adopted on the dataset with a small amount of data,which not only enables the network to have more data amount during training,but also is a more effective evaluation method to further improve the accuracy of Person Reidentification during testing.