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

With the improvement of the safety awareness of the whole society,video surveillance technology is widely used in public places to ensure property safety and personal safety.However,manual screening of target pedestrians in surveillance videos consumes manpower and time,and the negligence of supervisors may also lead to the omission of screening of target pedestrians.Therefore,the research of intelligent video surveillance technology is of great significance.In intelligent video surveillance,in order to realize the identification and tracking of the target pedestrian,pedestrian detection and pedestrian re-identification technology play a key role.Pedestrian detection locates and classifies pedestrian targets in the image,while pedestrian re-identification solves the identification of pedestrian identities under cross-cameras.In order to pursue the best recognition and positioning performance,the current detection model often has a huge amount of parameters and calculations,and the detection speed of the model cannot meet the needs of the application.At the same time,due to the existence of pedestrian pose changes,illumination changes,and occlusion,the pedestrian representation extracted by the network cannot be accurately identified,which brings a huge challenge to pedestrian re-identification.Aiming at the problems of pedestrian detection and pedestrian re-recognition,this article aims to study the intelligent monitoring technology in actual scenarios,propose a lightweight pedestrian detection model and fine-grained pedestrian re-recognition model,and apply them effectively in substation scenarios.The main research work of this paper is as follows:(1)A lightweight real-time pedestrian detection network is proposed,which realizes the balance between detection speed and accuracy.The model optimizes the network structure based on YOLOv4-Tiny,uses two different scale feature maps for target prediction,and improves the detection performance of the model from three aspects.Firstly,a multi branch structure embedded in the attention module is used to focus on the channel and spatial response at the same time,and the features of different receptive field sizes are integrated to obtain rich semantic information.Secondly,group self-attention is designed to mine context information to ensure that the amount of calculation and parameters of the model will not significantly increase while improving the detection accuracy.Finally,a hierarchical feature fusion network is introduced to make full use of the feature maps in different stages of the network to effectively fuse high-level features and low-level features to obtain fine-grained features.(2)A person re-recognition algorithm based on local features and attention mechanism is proposed to enhance the feature extraction ability of the model and suppress the interference of background noise.The algorithm takes Res Net-50 as the backbone network and uses the dual flow structure to extract global and local features respectively,so as to improve the recognition performance of the model from two aspects.Firstly,a multi branch channel aggregation module is embedded in the backbone network to mine the cross-channel correlation of different branch features,and effectively fuse the features of different receptive fields to obtain fine-grained pedestrian representation.Secondly,the cross regional attention module is introduced to make any body part feature contain the feature information of other body parts,enhance the richness of local features,and effectively resist the interference caused by pedestrian posture change,occlusion and other problems.(3)The local dataset is constructed based on the surveillance video taken by one camera outside the main control room of the substation and two cameras on the transformer in the park.The stability and robustness of the lightweight pedestrian detection and fine-grained person re-recognition algorithm proposed in this paper are verified on the local dataset.Secondly,combined with pedestrian detection,person re-recognition algorithm and target tracking algorithm,pedestrian tracking under single-camera and multi-camera in substation scene is realized,the recognition and tracking ability of target characters in wide-area scene is improved,and the intelligent monitoring and intelligent security level of substation are enhanced.