Research On Compression Method Of Deep Neural Network Model Oriented To Person Re-identification

As the main body of many places and events,people are the key tracking and identification objects of intelligent video surveillance in various places.Person Re-Identification（ReID）has become an important technology in intelligent monitoring systems.However,the widespread application of personnel re-identification technology to practical scenarios faces many challenges.Complex neural networks often need to take up a huge amount of memory space,and a huge amount of parameters will reduce the speed of re-recognition,and it is even more difficult to achieve real-time personnel re-recognition tasks.On the other hand,in reality applications,people moving in monitoring are always blocked by surrounding objects,which increases the difficulty of re-identification of people.Therefore,in order to make the person re-identification technology more widely used in real life,this article has carried out the following research on the problems of large model storage space and personnel image occlusion.（1）Aiming at the problem of large model memory and many parameters,this paper proposes a lightweight human body key point detection network.The LightBlock module is proposed by improving the core module of the MobileNet V2 network,Bottleneck,and the key point detection model LRP-Net is constructed.In the training phase,the key point heat map technology is used to obtain the personnel coordinates and offset,and the key point coordinates of the personnel are obtained through the regression algorithm,which greatly reduces the amount of network calculations.（2）Aiming at the occlusion problem in the process of re-recognition of persons,this paper proposes a re-recognition algorithm based on human high-order semantic features.Combining LRP-Net to obtain the key point information of the human body,and using the proposed person re-identification method LH-ReID to realize the learning of high-order semantic feature information of the human body.In the feature alignment stage,the affinity matrix and PReLU activation function are introduced to complete the matching of high-order key point information.The experimental results show that the model proposed in this paper occupies a smaller memory space and has a higher re-recognition accuracy rate than other models.（3）The proposed lightweight personnel re-identification model LH-ReID is applied to the logistics field,and the task of re-identification of the staff at the station is realized.LH-ReID and multiple common personnel re-identification models are applied to re-identify station employees,and Rank-1,mAP,GFLOPs,FPS and parameters are selected as performance evaluation indicators.Experiments show that the lightweight LH-ReID greatly reduces the amount of parameters and calculations of the model,speeds up the calculation of the network,and can better adapt to real-time personnel re-identification tasks.