The Implement Of Lightweight Person Re-Identification Algorithm

With the rapid development of modern society,more and more people pay attention to public safety.In the field of public security,video surveillance is widely used as the main means.In my country,the video surveillance system has been rolled out on a large scale,providing a strong guarantee for managing urban order and maintaining social stability.As an image search technology,pedestrian re-identification can play a key role in video security.Although with the rapid development of deep learning technology,pedestrian re-identification technology has made very significant progress.However,in many edge scenarios,the implementation of algorithms is still a problem that needs to be solved.The main contents of this paper include the following points:1.For the needs of lightweight networks,a lightweight backbone network basic component IBB-block is proposed in combination with the depthwise separable convolution structure and the IBN module,and the CBAM attention mechanism module is introduced to improve network performance.2.Choose to use distillation to improve the performance of the lightweight network,and in order to make better use of the intermediate feature information of the teacher network,use the feature distillation method for network distillation.In the experimental part,the effectiveness of the distillation method was verified through the ablation experiment,and in the subsequent distillation experiment,the effectiveness of the feature distillation method was illustrated by comparing the output distillation method.3.Design a set of unlabeled data collection process combined with existing mature algorithms,and then process multiple videos through this process to obtain unlabeled data.In the distillation process,the collected unlabeled data is used for the training of the lightweight network to improve the generalization ability of the network,and in the experimental part,the effectiveness of the method is verified by the leave-one-out verification method.4.According to the needs of the application,two deployment methods are proposed for the embedded deployment of the model.One is to deploy perceptual quantization on ordinary networks,and the other is to directly deploy lightweight networks through inference engine optimization.The deployment of the above two networks has been completed on the Jetson Xavier NX platform.The ordinary network has greatly reduced computing resources and memory requirements through perceptual quantization methods.While the accuracy of the network has slightly decreased,the inference speed of the network has been greatly improved.5.A pedestrian re-identification system was designed for the embedded platform,which realized the functions of pedestrian video analysis and pedestrian image query.In the system test,the usability of the system is verified.