Research On Person Re-identification Of Multi-camera Video Targets Under Deep Learning Conditions

The construction of new smart cities across the country is surging,and in the future the cities will be the all-connected cities.Monitoring equipment must be in every important corner of the city,and there is a huge amount of surveillance video data.Therefore,there is an urgent need for intelligent systems to label,process,and analyze the massive video data.Person re-identification technology is one of the core technologies in the fields of intelligent video surveillance,criminal investigation and security.It includes technologies such as data preparation,network construction and training,feature extraction and distance measurement.Focusing on the topic of multi-camera person re-identification of video targets under deep learning conditions,this paper carries out relevant work in four aspects: data intelligent annotation,fine segmentation of video target region,person re-identification based on image and mask,and person re-identification based on video sequences.The main work of this paper is as follows:1)Aiming at the primitive data annotation method,single function and complicated operation and the inefficiency of dataset production,this paper proposes a dataset labeling method for enhancing deep learning based on the intelligent annotation concept of human being in the ring.On this basis,an interactive data automatic annotation verification system based on deep learning is designed.The experimental results show that the method integrates automatic labeling,correcting errors,updating datasets,strengthening training,updating models and re-labeling functions.And the method improves the efficiency and accuracy of image and video data annotation,which is more conducive to the extraction of video target region in the bounding boxes.2)Aiming at the background interference in the bounding boxes output by target detection algorithms of the deep learning,this paper analyzes the commonly used video target region segmentation methods in order to improve the accuracy of person re-identification.On this basis,a fine segmentation algorithm for target regions in the bounding boxes under deep learning conditions is proposed.Firstly,using the background pixel information around the rectangular frame where the target region is located,the sub-block region growing algorithm is used for coarse segmentation,and then SRG Grab Cut algorithm is used for fine segmentation.The experimental results show that the algorithm can extract the target region in the bounding boxes quickly and effectively,which is more conducive to the subsequent person re-identification.Finally,a fine segmentation algorithm verification system under deep learning conditions is designed and implemented.3)Aiming at the difficulty that different person features are similar and the same person features are not the same in the current person re-identification task,this paper proposes an improved method based on image and mask for person re-identification.Firstly,based on the Resnet50 framework and PCB-6 block ideas,two improved network structures are constructed.After the training is completed,the test is completed on the constructed image and mask dataset of testing.The experimental results show that the method can improve the accuracy of person re-identification.Finally,a person re-identification verification system for multi-camera panoramic monitoring under deep learning is designed and implemented.4)The information of single-frame person image is very limited.This paper analyzes the features of structured person sequences.Then the paper proposes a structured person sequence re-identification method based on DTWT algorithm.Firstly,based on the Resnet50 framework,the network model is constructed and trained on the MARS dataset.After the training is completed,the test samples are sent to the feature extraction network to obtain the feature matrix of the sample sequence,and then the distances between the query sequences and the sequences to be matched are calculated and sorted according to the proposed DTWT algorithm.The experimental results show that this method can directly explain the validity of multi-frame information and improve the accuracy of person re-identification by utilizing spatio-temporal information such as the motion between frames.