Research On Pedestrian Re-identification Based On Deep Neural Network And Attention Mechanism

Person Re-identification is the task of searching pedestrian across multiple surveillance cameras.Person Re-identification plays an important role in the management of urban order,the tracing of missing persons and the detection of cases.Because the pedestrian appearance is greatly changed by the change of shooting angle,light and shade,pedestrian posture,and the structure of the upper pedestrian picture is relatively single,the difference between different types of pictures is mainly reflected in the fine-grained information of clothing and human body parts.Therefore,the first key to solve the problem of Person recognition is to obtain discriminatory information from the cluttered pedestrian recognition pictures.On the other hand,target detection algorithm is usually used to automatically extract pedestrians from pedestrian pictures and then to input the pedestrian recognition network.However,due to the deviation of the target detection algorithm and the different pedestrian postures in different pictures,the pedestrian features extracted from different pictures can not be aligned,which is the second key problem of pedestrian recognition.In order to solve these two problems,we try two different solutions:1)The first method uses the combination of global features and local features.We learn global features through network,segment global features horizontally to get horizontal local features,and segment global features vertically to get vertical local features,and then stitch these features together as the final expression of the input image.2)Attention mechanism is added to pedestrian recognition network to learn the weight distribution of feature map.We test the neural network designed under two solutions on three popular open datasets Market-1501,DukeMTMC-re1D and CUHK03,and verify the improvement of our method on mAP,rank 1,rank 5 and rank 10 by ablation experiments.Compared with other existing methods,our method has certain advantages in mAP,rank 1,rank 5,rank 10 indicators compared with the recent mainstream methods.