Research On Key Technologies Of Person Re-identification Based On Deep Learning And Attribute Features

Person re-identification is an important branch of research in the field of computer vision.Its technology is widely used in urban traffic management,public security,criminal investigation and other fields that require intelligent video surveillance.The main task of person re-identification is to solve the matching problem of the same pedestrian in the view of surveillance cameras with non-overlapping perspectives.In recent years,with the continuous development of the field of computer vision and the widespread deployment of intelligent surveillance cameras,person re-identification technology has become a key research issue for scholars at home and abroad.This paper summarizes the current research status of person re-identification at home and abroad,and based on the main challenges faced by current research,proposes a person reidentification feature extraction method based on neural network intermediate layer features and a person re-identification network combining pedestrian attributes and identity information.The main work done in this paper is as follows:1)Aiming at the problem of the background changing of pedestrian images in different camera angles,this paper proposes a background suppression algorithm based on PSPNet and kernel function weighting.PSPNet as an image semantic segmentation network,is used to preliminarily filter the background of pedestrian images.And then we weight the horizontal direction of pedestrian images by a kernel function,which is used to compensate pedestrians' horizontal pixel instability caused by walking.2)Research on pedestrian re-recognition methods based on deep neural networks and the fusion of multi-fine-grained features of mid-level information and high-level semantic features of neural networks.Verified by comparative experiments,the network model Residual network(Res Net)with the best performance in the pedestrian re-identification task is obtained,and based on this network,a fusion pedestrian feature description combining the mid-level information of the middle layer of the network and the high-level high-level semantic information is proposed.3)Research on the person re-identification algorithm based on the enhancement of pedestrian attribute matching model to realize the fusion of pedestrian identity characteristics and pedestrian attribute characteristics.Based on APR,a multi-task learning network for pedestrian identity and attributes with excellent model performance,this paper studies the construction of pedestrian attribute matching models to reorder the original feature distance rankings to improve the recognition accuracy of person re-identification networks.Combined with the operation of splicing pedestrian attribute features and identity information features in the original APR network,the pedestrian attribute features are used from multiple scales to optimize the original pedestrian feature distance ranking.In order to verify the effectiveness of the method in this paper,three independent person re-identification public data sets VIPe R,Market1501 and Duke MTMC-re ID were used to verify the experiment.The experimental results show that,compared with the current main pedestrian re-recognition algorithm,the algorithm proposed in this paper performs well on three public data sets,and reflects a certain accuracy rate advantage in recognition rate.