Research On Person Re-identification Method For Security Scenes Based On YOLOv5 And DeepSORT

Person re-identification refers to identifying and detecting the identity of the same person from different perspectives,camera lenses,and angles.Currently,research on person re-identification mainly focuses on image retrieval,and there is a lack of direct person re-identification methods for monitoring videos.However,with the widespread use of video surveillance equipment and monitoring systems,there is a higher demand for the accuracy and real-time performance of re-identification methods.Therefore,pedestrian recognition technology that can meet the requirements of real-time,accuracy,and intelligence in real-life has significant practical application value in the security field.The application of network optimization technology can help improve the accuracy and efficiency of deep neural networks.This paper divides the person re-identification process into two stages: pedestrian detection and person re-identification and proposes a multi-objective vector representation method based on genetic algorithms and its optimization method.The optimized pedestrian detection model is used to detect all pedestrians in the video frame,and then the re-identification model is used to match pedestrians to achieve person re-identification in security monitoring videos.The main work of this paper is as follows.(1)In terms of network optimization,a neural network structure vector representation method based on deep learning is designed for the multi-objective optimization problem of deep neural networks.A multi-objective optimization model for neural networks is proposed,and a multi-objective optimization method based on genetic algorithms is implemented and tested on the MNIST dataset.The experimental results show that compared with a four-layer fully connected network set by traditional empirical settings,the network accuracy is increased by 1%,and compared with the network optimized by the Tensor Flow library’s optimization algorithm,the accuracy is increased by 10.41%.The overall number of network parameters is reduced by 32.6% compared with a four-layer fully connected network set by traditional empirical settings and by 13.2% compared with the network optimized by Tensor Flow’s optimization algorithm.The experiments verify that this algorithm can perform multi-objective optimization on deep neural networks,improve network accuracy,and reduce the overall number of network parameters.(2)In terms of pedestrian detection,to obtain a more suitable pedestrian detection model for this paper’s research environment,the proposed optimization algorithm is used to perform multi-objective optimization on YOLOv5’s m AP＿0.5 and m AP＿0.5:0.95 and tested on the COCO128 dataset.The experimental results show that the precision of the optimized model reaches 86.2%,the recall rate reaches 85.7%,the m AP@0.5 reaches 92.1%,and the m AP@0.5:0.95 reaches 59.2%.Compared with YOLOv5’s pre-training model YOLOv5 x,the optimized model improves the precision,recall rate,and m AP@0.5:0.95 by 5.9%,10.4%,and 8.8%,respectively,and reduces the m AP@0.5:0.95 by 2.5%.Through experimental analysis,it is verified that the proposed optimization algorithm can reduce the resources required for network training,improve network recognition accuracy and training efficiency,and provide new ideas for optimizing large-scale deep neural networks.(3)In person re-identification,to further improve the re-identification performance,this paper uses Fast Re ID to train a better feature extraction model,which replaces the original feature extraction model of DeepSORT.Fast Re ID’s feature extraction model is tested on the Market-1501 and Duke MTMC-re ID datasets.The experimental results show that on the Market-1501 dataset,compared with the bagtricks＿R101-ibn model,the sbs＿R50-ibn model reduces rank1 by 0.06% and increases m AP by 0.57%,without significantly affecting accuracy,but significantly reducing training time by 51.49%.On the Duke MTMC-re ID dataset,the sbs＿R50-ibn model increases rank1 and m AP by 0.9% and0.95%,respectively,and reduces training time by 44.31%,achieving a balance between accuracy and efficiency.Therefore,the Re ID model trained using the sbs＿R50-ibn network is adopted for replacement,and it is validated on the MOT16 dataset of surveillance videos,showing that it can effectively solve the ID switching problem in pedestrian tracking and achieve real-time tracking.(4)In practical applications,based on the security scenario,this paper uses the person re-identification algorithm based on YOLOv5 and DeepSORT to build a video surveillance intelligent analysis platform.The platform includes functions such as camera registration and deployment,pedestrian database and group management,offline video uploading,simulated cameras,video surveillance,and intelligent analysis result viewing.The platform is designed with a simple and practical visual interface,which greatly improves the efficiency of monitoring analysis,effectively saves the energy and time required for monitoring information analysis,and achieves good results in practical scenarios.