Research And Application Of Pedestrian Detection And Person Re-Identification Based Deep Learning

With the continuous development of society and accelerated urbanization,pedestrian flow is increasing and the demand for pedestrian detection and pedestrian recognition is growing.Pedestrian detection can play an important role in areas such as video surveillance and intelligent transportation,for example,to achieve functions such as personnel access management and traffic road supervision.Pedestrian re-identification,on the other hand,can be used to retrieve specific pedestrians in large-scale data or for applications such as pedestrian behavior analysis.Research on pedestrian detection and re-identification is important for improving social safety and traffic flow.In practical applications,there are still some challenges in pedestrian detection and rerecognition,such as the problem of data distribution bias,the models trained on public datasets cannot cope well with practical application scenarios,and the algorithm data processing efficiency cannot meet large-scale applications.The research in this thesis focuses on how to solve these difficulties and improve the robustness and efficiency of the algorithm to meet the needs of more practical application scenarios.At the same time,issues such as pedestrian privacy protection need to be considered to ensure that the technology application can achieve social acceptance and promotion.For the pedestrian detection problem,a lightweight pedestrian detection network based on the improved YOLOv5 s network is designed.The Conv structure in the original YOLOv5 s network is replaced by a more lightweight Ghost Module module to achieve the purpose of reducing the number of network parameters.At the same time,the GAM global attention mechanism is introduced in the Neck output part of the network,which effectively improves the target detection performance of the network.After experimental comparison,the network designed in this thesis achieves 73.6% m AP on the self-built dataset.For the pedestrian re-identification problem,a multi-branch pedestrian re-identification model incorporating pedestrian attribute information based on Res Net50 as the backbone network is designed.The network first optimizes Res Net50 by dividing the Stage4 structure of the last part of the network into two branches to reduce the interference between multi-task learning,and in order to improve the description of pedestrian detail information in the feature map,the downsampling operation in the Stage4 structure is eliminated so that the output feature map size becomes twice as large as the original one.Then the two branches of the network learn the global identity features of pedestrians and the attribute features of pedestrians,respectively,to improve the model performance by introducing the attribute information of pedestrians.Finally,experiments are designed to demonstrate the effectiveness of the network model.Finally,based on the above research on pedestrian detection and pedestrian re-identification,a campus pedestrian retrieval system is designed and constructed,including modules for pedestrian retrieval,real-time video viewing,pedestrian library viewing,and device management,etc.Finally,the system is tested to verify the use of the algorithms in this thesis in practical scenarios.