Research And Implementation Of Person Re-identification System Based On Deep Learning

Person re-identification aims at retrieving images of a specified pedestrian across the cameras,given a probe image.As one of the key intelligent video surveillance analytic technologies,person re-identification is an effective method to find a pedestrian target based on pedestrian features from the whole body on images,when the camera can not capture pedestrian face.In the face of massive video surveillance data,manual search is inefficient.Using person re-identification technology can simply and efficiently locate the person of interest,which is in effect to promote public safety.However,due to the dramatic feature variations caused by background,light and human pose on so on,extracting discriminative features has become a key point of person re-identification.In recent years,deep learning methods dominate this computer vision community,with a great superiority against traditional methods.Using deep learning methods to solve the pedestrian recognition task becomes mainstream.Therefore,this thesis makes an deeply study on the method of feature extraction based on deep learning,and implements the person re-identification system.The specific work of this thesis is listed as follows:1)In the study of loss function,in order to implementing the target that samples with the same identity are closer to each other than those with different identities,a novel loss named Triplet-Center Loss is proposed for person re-identification,which combines the advantages of both Center Loss and the Triplet Loss.Compared to other metric loss functions such as triple loss,It can effectively optimize the feature embedding space.In the view of the problem that Softmax Loss easily leads to the learned deep model over-fitting,label smooth strategy is used to improve the generalization ability of the model.Finally,the person re-identification is trained via Triplet-Center Loss and Label Smooth Softmax Loss.The experimental results demonstrate the effectiveness of the proposed method.2)From the perspective of neural network structure,the Multiple Granularity Network with attention mechanism is designed.The mixed attention mechanism combining the spatial and channel attention is proposed to reduce the impact of the environment variations and guide deep neural network to pay more attention to pedestrian targets.According to multiple image partition,The network can obtain more discriminative multi-granularity local features while focusing on global features.The whole network trained via Triplet-Center Loss and Label Smooth Softmax Loss shows the superiority in experiments.3)Based on the feature extraction model,the person re-identification system is designed and implemented.In addition,the function test and performance test of this person re-identification system is carried out.The results prove that the system meet the basic needs of users.