Research On Pedestrian Re-identification Algorithm Based On Deep Learning

With the the development of Smart City construction,the field of pedestrian detection and pedestrian re-identification has received extensive attention from the academic community.Pedestrian detection and re-identification have been widely used in Police Criminal Investigation,Smart Transportation,Epidemiological Survey and other fields.The increasing scale of surveillance in cities will produce a great deal of surveillance video data,which provides a basis for tracking specific targets and puts forward higher requirements for pedestrian detection and reidentification algorithms.However,there are still some problems in the application of pedestrian detection and reidentification due to the complex realistic environment,such as:(1)The real-time performance of pedestrian detection algorithm is poor and the boundary description is not accurate enough;(2)Reidentification datasets are too small to achieve uniform feature embedding and classification results,and the generalization ability of models across domains is insufficient;(3)The pedestrian re-identification task in real scenes has occlusion phenomenon,which causes many problems of wrong matching and feature confusion.Therefore,based on deep learning technology,this thesis has carried out the following researches in the field of pedestrian detection and re-identification.On the issue of problem(1),the text detection algorithm DBNet is improved,and a differentiable binary pedestrian detection algorithm(ASPP-DBNet)based on atrous spatial pyramid pooling is proposed.Its backbone network can keep a large receptive field while integrating multi-scale features,and adaptively learn boundary thresholds through differentiable binarymodule,effectively balances performance and efficiency of pedestrian detection.At the inference stage,a segment-candidate box transformation algorithm based on Bessel curve fitting is proposed,which can transform the segmented connected domain of network output into candidate box coordinates.According to the experiments,the FPS of lightweight ASPP-DBNet frame count is 61,while the F-1 and m IOU indices on the pedestrian detection dataset are 83.84% and 73.4%,respectively,better than DBNet and other mainstream algorithms.On the issue of problem(2)and(3),this thesis proposes a feature fusion pedestrian re-identification network based on pose estimation from two aspects: solving the occlusion problem and global-local feature fusion.The model uses key points of the human body to judge the blocking area and insert the branch of the feature pyramid.The feature is that the blocking area of the image can be measured,and then the visible local features can be merged with the global features to improve the representation of the network.In addition,attention mechanism and orthogonal regularization are introduced in the branching of feature pyramid to ensure feature diversity.The experiments are carried out on three full-body datasets and one occlusion dataset.The algorithm’s Rank-1 index on the full-body dataset is 1.54%,1.0%,and 2.67% higher than the suboptimal algorithm,respectively.On the occlusion dataset,the Rank-1increased by 1.38%.Using the proposed ASPP-DBNet pedestrian detection algorithm and the pedestrian re-identification algorithm based on pose estimation feature fusion,an abnormal person intelligent retrieval system is constructed.The system functions include front-end display interface,system management,pedestrian image database management,etc..Finally,the performance of the system is tested,and the results show that the algorithm in this thesis also has high reliability in practical application scenarios.The thesis includes 40 figures,22 tables,and 86 references.