| Person Re-identification(Re ID)is a technique that uses computer vision to find specific pedestrians in images or video sequences.In recent years,person re-identification technology based on deep learning has been widely used in practical fields such as intelligent security and video surveillance.In this paper,based on the deep learning algorithm,the person re-identification is deeply studied.There are three difficulties including the accuracy of the benchmark network,the network convergence time and the sufficiency of the extracted features.Aiming at three problems,this paper makes three improvements.The main work and the conclusions of this paper are summarized as follows:1.Aiming at the problem that the accuracy of the benchmark network is not high enough,this paper proposes a re-ID algorithm based on powerful batch normalization and attention adaptive fusion.Firstly,a powerful batch normalization module is used to jointly combine the metric learning loss function and the representation learning loss function for network training.Secondly,this paper adds an attention adaptive module to the benchmark network,which encourages the model to adaptively pay attention to the key regions of the image and select valuable information for person reidentification,thereby improving the accuracy of the benchmark network.2.Aiming at the problem of slow network convergence,this paper proposes a re-ID algorithm based on siamese network.Firstly,siamese network includes the same backbone network with shared weights,which enables the network to extract features from two images simultaneously.In addition,the weight sharing of the network can also speed up the overall convergence speed.Secondly,this paper also adopts the intelligent optimization sorting module,which greatly improves the recognition accuracy.The corresponding experiments in this paper demonstrate the superiority of using the siamese network and the intelligent optimization sorting module,and the recognition accuracy(Rank-1)and Mean Average Precision(mAP)indicators on different datasets have also been improved.3.Aiming at the problem that the extracted features are not sufficient,this paper proposes a reID algorithm based on multi-branch feature fusion structure.Firstly,this paper uses a heat map to show which parts of the input image play a major role in the final classification judgment,and comprehensively analyzes the advantages and disadvantages of different feature extraction methods.Secondly,this paper adopts a multi-branch feature fusion structure to extract global features and local features of images respectively.In this paper,the horizontal segmentation method is used to extract local features,which can be combined with global features to obtain rich high-level semantic features.In addition,this paper adds a random erasure module,adding images of randomly erasing blank areas into the training data to enhance the anti-interference ability of the model.The visualization results of the experiments in this paper demonstrate the superiority of the multi-branch feature fusion structure. |
