Research And Application Of Person Re-identification Method Based On Deep Learning

Person re-identification is one of the important technological means in the fields of security monitoring,search and rescue,and commercial services for future smart cities.Accurately extracting person features is crucial to improve the security and management efficiency of smart cities.However,existing person re-identification methods still face challenges such as susceptibility to lighting conditions,large model parameters,insufficient accuracy,and low level of intelligence.Moreover,traditional person feature extraction methods have certain limitations and cannot meet practical application requirements.Therefore,it is of great significance to conduct research on deep learning-based person re-identification methods to achieve accurate extraction and calculation of person features and improve the efficiency of person re-identification.This paper conducts in-depth research on this topic,and the main research works are as follows:(1)In response to the problems of blurred details,low contrast,and difficult information extraction in video images captured by surveillance equipment in low-light environments,a low-light image enhancement method that integrates multiple-scale dense networks is proposed.By synthesizing different illumination datasets to simulate real nighttime scenes,an attention mechanism is introduced to characterize the correlation between different channels and highfrequency information.Then,a multi-scale fusion residual dense connection network is constructed to enhance overall detail perception.Finally,a bilinear additive convolution structure is used to replace the deconvolution and eliminate the ghosting phenomenon.Experimental results show that compared with mainstream methods,this method improves PSNR and SSIM by 26.37% and 14.14%,respectively,and the enhancement effect is significant.(2)In response to the problems of excessive parameters and insufficient accuracy of lightweight models in existing person detection methods,a lightweight person detection method based on improved YOLOv5 s is proposed.The method uses a lightweight C3 Ghost module to replace the main network C3 module to reduce the parameter volume,introduces an efficient channel attention mechanism to fully extract features,and uses the lightweight convolution method GSConv in the neck network to reduce the computational cost of person detection.The SIo U loss function is used to accelerate the convergence of the training process to improve the accuracy and efficiency of the model.Experimental results show that the proposed method improves the average m AP by 3.79%,reduces the model parameter volume by 22.85%,and reduces the floating-point operation volume by 24.52%,which verifies the effectiveness of the improved method.(3)In response to the problem of insufficient feature extraction of person characteristics in current person re-identification models leading to insufficient model accuracy,a multi-level feature fusion person re-identification method based on improved BDB-Net is proposed.The method introduces a new feature deletion strategy to enhance the network’s attention to low activation features,adds a regularization branch to reduce noise interference caused by feature deletion,and introduces a Res Ne St-50 backbone network and Split attention mechanism to improve attention to global human body features.Experimental results show that the proposed method achieves good performance with an average precision of 88.5 and a Rank-1 index of95.8 on the Market1501 dataset.Ablation experiments demonstrate the effectiveness of the improvement strategy.(4)Based on the research results of the paper and combined with the intelligent application requirements of real-world scenarios,a person retrieval system is designed and constructed.The system includes a graphical user interface,low-light image enhancement,and person retrieval function combined with person object detection and person re-identification technology.The test results of the system show that the proposed method has high practical value and significance.