Research On Person Re-identification Method Based On Deep Learning

In recent years,the development of urban intelligence has become more and more mature,and all major cities have deployed a large number of high-definition surveillance cameras,thus acquiring a large amount of video information.Among them,intelligent analysis and utilization of pedestrian data has become an important research topic,while person re-identification refers to the use of computer vision techniques to determine the identity of a specific person in multiple non-overlapping area surveillance camera systems and can be considered as a sub-problem of image retrieval.This paper discusses and investigates the deep learning-based person re-identification technology,which is mainly as follows:(1)Lightweight person re-identification algorithm with multi-scale fusionAlthough the current supervised learning-based person re-identification network has a high accuracy rate,there are problems of large hardware resource requirements and low operational efficiency in practical applications.Therefore,in this paper,we design a lightweight person re-identification network model,which uses four convolutional blocks to extract features at different scales from the bottom detail information to the top semantic information of the image step by step,and then fuse the feature information at different scales after pooling operation.The fused features are then input into the global attention module,and the feature information with poor recognition ability is filtered out by setting different weighting parameters to optimize the fused features.The results show that the number of parameters is reduced by more than half compared to the mainstream network architecture,and the recognition accuracy is higher than that of traditional benchmark methods.(2)Unsupervised person re-identification algorithm by part-compensated soft multi-label learningSupervised learning-based person re-identification algorithms require a large amount of manual data labeling and are less scalable in practical applications.To address the problems such as reduced recognition ability of the algorithm due to the lack of data annotation,this paper proposes an unsupervised method for learning locally compensated soft multi-labeling,which mainly consists of global clustering,local clustering and multi-label assignment.In the clustering process,the results of local clustering are compensated to the global clustering so that the model can fully learn more local feature information of pedestrians.Meanwhile,soft multi-labeling is generated using fuzzy clustering to solve the problem that a large number of positive samples are discarded,and cross-domain excitation parameters are set to solve problems such as background interference in inter-domain views.The results show that the method outperforms the currently common unsupervised learning network models.(3)Semi-supervised person re-identification algorithm based on human pose augmented soft label learningAlthough the unsupervised learning-based person re-identification algorithm can solve the problem of lack of data labeling,the recognition accuracy is not satisfactory because it completely discards label information.Therefore,this paper uses a small amount of label information to optimize unsupervised learning and designs a semi-supervised learning based person re-identification network with soft label learning using human pose enhancement.First train an accurate pedestrian pose guided occlusion predictor using labeled data,which can effectively shield the effects from occlusion during unlabeled training.Then,after combining the predictors to generate accurate feature vectors,the feature affinity matrix is calculated and soft label information is generated to optimize the network.Experiments were conducted on three public datasets,Market-1501,Duke MTMC-re ID and MSMT17,and the results demonstrated the validity and reliability of the method.