High-performance Person Retrieval And Re-identification Technology

Person search algorithms are widely used in intelligent security systems in large public places,and can be decomposed into two tasks: pedestrian detection and pedestrian reidentification.In the complex cross-camera real scene,the person detection algorithm is used to obtain the region of interest(ROI)of a person first,and then the feature vector is used to obtain the feature vector of the person ROI image,and the person matching algorithm is used to determine whether the target person appears in other cameras with cross-camera recognition.This thesis mainly focuses on improving the person detection speed and person re-identification accuracy through theoretical research and a large number of experiments.Based on the Sophon SC5+ artificial intelligence chip,the improved algorithm is quantified,and a high-performance person search system is deployed to real-world scenarios.The main work of this thesis is as follows:1.Design a lightweight person detection algorithm.This paper proposes to use the Ghost Net network to replace the computationally expensive CBS module and C3 module in YOLOv5,which greatly reduces the amount of model parameters.However,the reduction of the number of parameters results in the decrease of accuracy.To alleviate this issue,the CBAM attention mechanism module is used to adjust the features in the two dimensions of channel and space to further strengthen the feature extraction ability of the network.Finally,in order to meet the requirements of real-world applications,the improved YOLOv5s-GS is trained on a dedicated person detection dataset,making the trained model more suitable for intensive person detection.The experimental results show that the improved algorithm in this paper reduces the amount of model parameters and improves the detection speed under the condition that the accuracy is nearly unchanged.2.A person re-identification algorithm with fusion loss optimization is proposed.In order to solve the problem of fuzzy optimization target in the triple loss of hard sample mining,two positive sample pairs and one negative sample pair are used to improve the network,so that the network can learn more comprehensive features.In addition,in order to further improve the generalization ability of the trained network,the intra-class feature distance is compressed and the center loss constraint is added.The improved centerconstrained triplet loss and additional interval cross-entropy loss are jointly combined for model-supervised optimization,which combines the advantages of metric-based learning loss and representation-based learning loss.Finally,the BNNeck network structure and Non-Local attention are used to solve the problem that the fusion loss function is difficult to converge during training.3.Combine the improved lightweight person detection model with the person reidentification model,and deploy the model based on the Sophon SC5+ artificial intelligence chip.In order to verify the high performance of the quantized model,the algorithm inference result under 2080 Ti is compared.According to the image processing and hardware acceleration interface provided by the Sophon platform,the person search core interface is designed and compiled.Through the verification of the running results of the algorithm,the top-down end-to-end optimization method,i.e.,optimizing from the algorithm to the hardware,reachs an optimal efficiency of deep learning.The experimental results show that the person search system based on Sophon has low cost and high speed,and can complete the rapid retrieval of target persons in surveillance videos.