Research On Person Re-identification Based On Deep Learning

As society continues to develop,people are paying more and more attention to public safety issues.Nowadays,surveillance cameras are arranged in public places such as schools,shopping malls,streets,stations,etc.Video surveillance systems play an important role in managing urban order and maintaining social stability.Person re-identification,as a technology to retrieve specific pedestrians across cameras,is a key technology in intelligent security and has become a hot topic of research in recent years.However,due to the complexity of the actual shooting scene,the acquired pedestrian images have some problems,such as posture change,occlusion,low resolution and so on.How to extract distinguishing pedestrian features has become a difficult task for person re-identification.Deep learning method can extract the feature information of the pedestrian image,more distinguished and robust than the traditional method and robustness,and can implement end-to-end training.Therefore,this paper investigates the task of person re-identification in conjunction with deep learning algorithms,with the following main work and contributions.(1)To address the limitations of local feature learning and the underutilization of lowlevel features,Multi-scale and Multi-granularity Fusion Network(MMF-Net)is proposed,aiming to utilize feature information more comprehensively and effectively.MMF-Net uses a multi-branch structure to learn features at different scales and different granularities,and optimizes global features with local feature learning,which strengthens the global and local correlation.Meanwhile,the semantic supervision module is introduced in the lower layer of the network to extract low-level features,and it is used as a complement to the pedestrian image similarity metric to take advantage of the complementary advantages of low-level and high-level features.In addition,the improved pooling layer in this paper combines the features of average pooling and max pooling to obtain more discriminative features.(2)Aiming at the problems that the existing person re-identification methods focus too much on the extraction of strong discriminative features of pedestrian images leading to the lack of robustness of the model and cannot well combine spatial and channel dimension information,a person re-identification method based on spatial weakening and channel enhancing attention is proposed.By weakening the attention to the high response region,the method forces the model to learn more comprehensive feature information,which improves the generalization ability of the model.At the same time,a multi-branch structure is formed by combining spatially weakened modules at different layers of the network to increase the diversity of model feature representation.In addition,a channel attention mechanism is embedded in the network to learn correlations between feature channels and automatically calibrate attention on the channel dimensions.This paper performs a wide range of experiments on the four mainstream datasets of Market-1501,Duke MTMC-Re ID,CUHK03,and MSMT17 to verify the performance of the model and compare with excellent algorithms in recent years.The experimental results show that the proposed method has good robustness and recognition accuracy,and can effectively improve the performance of pedestrian re-identification.