Research On Person Re-identification Algorithm In Multi-camera Video Surveillance Scene

Person re-identification(Re ID)mainly studies how to find the same object of interest in multi-camera network.As one of the key technologies of video surveillance,it has high application value in public security fields,such as long-term object tracking,analysis of pedestrian abnormal behavior and so on,and is of great significance to realize intelligent security.In the multi camera video surveillance scene,the changes of pedestrian’s position,posture,lighting conditions,background interference,occlusion,camera parameters and other factors lead to the great apparent differences in the images of the same pedestrian in different camera viewpoints.Therefore,the challenge of person re-identification is two parts: to extract the pedestrian characteristics with higher discriminability and to design a distance metric algorithm with higher accuracy.To overcome the challenge of extracting strong discriminant features,a person re-identification method based on much attention activation mechanism is designed,the method exerts attention mechanism on feature maps in more than one abstract semantic stage,and attention mechanism in every stage uses multi-branch structure to extract significant pedestrian characteristics of multi layers and multi areas.What’s more,the model is designed as modular in order to improve the whole computational efficiency.To solve the problem of designing a more accurate feature distance measurement algorithm,a distance measurement algorithm for person re-identification based on graph convolutional network is designed.A subgraph is constructed for each query image,and is input into graph convolutional network for reasoning calculation to predict the feature distance between the query image and the neighboring image at all levels.The multi-attention activation method can extract strong discriminative features,while the graph convolutional network is good at measuring the distance between pedestrian features,so a method based on both attention mechanism and graph convolutional network is proposed,which combines the advantages of the first two methods.The multi attention activation mechanism is applied to the feature extraction stage,and the graph convolutional network is applied to the distance measurement phase of characteristics,in order to obtain the ability of strong discriminant feature extraction and the more accurate distance measurement.Ablation experiments on Market-1501、DukeMTMC-reID、CUHK03 and MSMT17 show that the three methods designed in this paper can effectively improve the accuracy of person re-identification.Among them,the method with the combination of attention mechanism and graph convolutional network achieves the best performance.The accuracy of Rank-1 and mAP on Market-1501 can reach 95.2% and 87.3%.