Research Of Person Re-identification Based On Faster-RCNN

In recent years,person re-identification technology has become a research hotspot in computer vision due to its great potential in public security and other fields.Essentially,person re-identification is an image retrieval process.With the extensive application of deep learning technology in image processing,convolutional neural network and its derivative network have become the mainstream basic technology to solve the problem of person re-identification.Aiming at the difficult problems that need to be solved urgently in the process of person re-identification such as unlabeled data and insufficient discrimination of pedestrian features,researchers have proposed many high-quality algorithm frameworks,and used common data sets to verify the algorithm,which has achieved remarkable results.On the basis of sorting out and summarizing the related theories of in-depth learning technology,elaborating and analyzing the development process,classical solutions and difficulties of person re-identification,this paper carries out innovative research in two aspects.Firstly,aiming at the problem of insufficient discrimination of pedestrian characteristics in the process of person re-identification using progressive unsupervised model,a centralized clustering loss function is proposed.Secondly,the person reidentification framework E2 E based on Faster-RCNN is improved,and the residual neural network ResNet-50 is used to extract pedestrian characteristics and optimize the model training process.To solve the problem of insufficient distinguishability of pedestrian features extracted after training with Softmax loss function in progressive unsupervised learning model,this paper sets distance penalty term and proposes centralized clustering loss function.In training,the pedestrian characteristics of the same identity are closer to the clustering center selected by k-means++,which enhances the distinguishability of pedestrian characteristics.A comparative experiment is designed and the pedestrian characteristics are visualized during the training process,which proves the effectiveness of the proposed method.Aiming at the problem that the data set of person re-identification is quite different from the actual application scenario,in the end-to-end person re-identification framework E2 E,ResNet-50 is used to replace AlexNet in E2 E,so that more detailed semantic features can be extracted from deeper network,target candidate regions can be generated by regionproposed network,and pedestrian detection and recognition can be realized through joint training of multiple loss functions.Integration of different processes.In the experiment,the proposed algorithm is compared with several traditional pedestrian detection algorithms and the original E2 E framework.The rationality and effectiveness of the proposed algorithm are proved.There are 42 images,7 tables and 63 references in this paper.