Research And Application Of High Precision Pedestrian Recognition In Complex Scenes

Person re-identification is a key technology for video intelligence analysis.Its core function is to judge the consistency of pedestrian identity in the case of cross-camera based on the extracted person features.At present,the focus of research on pedestrian re-identification lies in the extraction of person features.The method of manually extracting person features is mainly based on the low-level features such as color and texture.Recently,with the breakthrough progress of deep learning technology in the field of computer vision,the researches on person re-identification based on deep learning has become the research emphasis.When faced with complex scenarios.It mainly uses the complex neural network structure to extract the person features with strong robustness for person feature comparison.For complex scenarios,designing a simple and effective neural network structure to obtain the person features with judgment is the research focus of person re-identification in this paper.The main research of this paper has the following aspects:1)Analyze the influence of different loss functions on the accuracy of pedestrian re-identification model:Firstly,a simple classification task is designed to test several mainstream feature extraction networks.The experimental results show that the model based on Resnet50 feature extraction network get the highest accuracy.After selecting the Resnet50 network structure,we used the network structure of Resnet50 to perform several sets of experiments on the loss function of the classification task?the loss function of the verification task and combination the loss function of the classification task and the loss function of the validation task.The experimental results show that there is a certain complementarity between the appropriate classification task loss function and appropriate validation task loss.Reasonable combination them can improve the accuracy of the model.2)Propose a graph-based global reasoning network module:In the person re-identification task,global modeling and reasoning between different body region relationships is very necessary.Convolutional neural networks can only acquire local features by using convolution operations.Inspired by the attention mechanism,this paper proposes a new global reasoning method.The algorithm first clusters the feature map on the coordinate space then project it into an interactive space where relational reasoning can be effectively computed.After reasoning,relation-aware features are distributed back to the original coordinate space.Through theoretical analysis and experimental verification,the global reasoning network module is embedded in the original network to improve the network's ability to express features.3)Construct an application system with person re-identification service as the core:We have independently built a high-performance server for algorithm training,and completely deploys the person re-identification network service on the server.