Research On Person Re-Identification Based On Deep Learning

In current social life,information security and intelligent security are increasingly highlighting its value,and are important means of maintaining long-term security in society.With the increase in the number of surveillance cameras in public places,large-scale distributed video surveillance systems cover more and more areas in cities.Faced with massive video surveillance data,traditional manual video analysis methods have encountered bottlenecks and intelligent surveillance video analysis systems There is also an urgent need for optimization and upgrading.In order to better solve these problems,Person re-identification technology emerged at the historic moment,and became a hot spot in the field of intelligent security in a short time.With the rapid development of deep learning in the field of artificial intelligence,person re-identification technology has gradually changed from the traditional manual hand-craft methods to deep learning methods for research.Aiming at the problems of pedestrian poses variation and frequent occlusions in real life scenarios,this thesis uses deep learning methods to explore the problem of person re-identification and proposes an effective solution.The main work content and innovations are as follows:1.A local feature extraction method based on hard attention mechanism is proposed.This method can extract more detailed features of human body parts from the input pedestrian images.First,the pedestrian is positioned at joint points by a pose estimation tool,and the pedestrian body is divided into local areas.Then the hard attention module is applied to the local features.Through the hard attention learning,the regions with potential discriminative power can be located on the local feature map of the human body,and the features of human body parts with more fine contours are extracted.This method can eliminate irrelevant background noise interference,and at the same time alleviate the problem of occlusion between adjacent parts of the human body to a certain extent.2.A global feature extraction method based on soft attention mechanism is proposed.This method can separate the foreground and background of the pedestrian image,eliminate irrelevant background noise interference,and simultaneously extract pixel-level global features.This method uses convolutional neural network to extract pedestrian global features,and introduces a soft attention module to re-correct feature representations.The soft attention module consists of spatial attention and channel attention.In particular,this thesis models spatial attention and channel attention simultaneously in parallel,and compares it with a variety of serial structure methods.This kind of soft attention learning aims to select important pixels with fine granularity,and concentrate the network's attention on pedestrians as much as possible,eliminating the background noise interference.3.In summary,this thesis proposes an efficient and lightweight feature fusion network based on the hybrid attention mechanism.This network fuses local and global features to form a pedestrian descriptor,and uses this to distinguish pedestrians' identities.In this thesis,three large public data sets,Market-1501,DukeMTMC-ReID,and CUHK03-NP,are compared experimentally with a large number of state-of-the-art methods for person re-identification,and the reasons for the experimental results are analyzed in detail.More than that,this thesis analyzes the ablation experiments of each component of the feature fusion network.Extensive experiments and a large amount of experimental data prove the robustness and efficiency of the network model proposed in this thesis.