Person Re-identification Based On Attention Mechanism

Recently,with the rapid development of deep neural networks and the ever-increasing demand for intelligent video surveillance,surveillance networks composed of cameras have covered various places in city streets.The video surveillance systems are an important aid to the prevention and control of urban security.When a crime occurs,it is of great significance to track and retrieve the target pedestrian using the numerous pedestrian images collected from different cameras.Pedestrian recognition technology has therefore become one of the indispensable technologies in video surveillance systems,which is undoubtedly a challenging academic research.Using computer vision technology to determine whether the pedestrian images under different cameras are the same person involves how to extract recognizable features from the pedestrian images and design a better distance measurement method.However,in real scenes,pedestrian re-recognition technology still faces many challenges,such as lighting,obstacle occlusion,different resolutions of pedestrian images,differences in pedestrian postures,and different camera angles.In view of its research influence and the importance of practice,the research of pedestrian re-identification has received extensive attention from the academic community.For the above-mentioned problems encountered in the pedestrian recognition technology again,we propose the following two methods:1.Aiming at the problem of pedestrian image posture difference and pedestrian image misalignment,a pedestrian re-recognition algorithm based on posture guidance and feature alignment is proposed.Use Openpose to locate the human body posture point information,integrate the human body posture information into the global feature representation,use the hard attention mechanism based on the human body key points to force the network to pay attention to the pedestrian posture features;and use the human body key points to achieve the same body between the pedestrian images Align the positions to realize the task of re-identification of pedestrians.2.Aiming at the background noise in the pedestrian image will affect the problem of the network paying too much attention to useless information.A pedestrian re-recognition algorithm based on attention mechanism and multi-scale feature fusion is proposed.The algorithm proposed in Chapter 4 uses the last layer of features to collaborate with multiple middle layer features,and uses a top-down progressive summation pyramid feature fusion mechanism to extract pedestrian image features,using maximum pooling and average pooling,respectively The saliency information in the image is retained,and the attention module is added to the network to ensure that the experimental model improves the characterization ability of detailed information on the basis of the overall feature expression.The proposed two pedestrian recognition algorithm again,an effective solution to the pedestrian image is not overly concerned with useless network alignment and picture information and other issues,and then to identify the experimental data after tests showed that the three main data sets of pedestrians,mentioned The two algorithms have certain advantages.