Person Re-identification Based On Robust Feature Extraction And Network Optimization Compression In Open-world

Under the background of big data,the research on person re-identification(re-ID)based on statistics and video data has made great progress,and the output of research results has achieved high accuracy on public data sets.However,the performance of re-ID applications in the real world is very poor,mainly due to the single research variable,unreliable feature selection,too many network parameters,etc.In view of the above problems,the following research work has been done:(1)Aiming at the problem of single research variable and unreliable feature selection,a solution to extract pedestrian periodic motion characteristics in open-world to identify pedestrians is proposed.Specifically,the temporal and spatial motion features of pedestrians are extracted from the surveillance video based on the attitude estimation network,and in order to improve the ability of feature extraction of the attitude estimation model,the residual block is constructed based on block convolution to improve the network.Compared with the previous network,the improved network has 1.8% performance improvement,and saves 5% computing space and 15% storage space.At the same time,aiming at the defects of the existing re-ID data set,a test set containing factors such as pedestrian changing and cross-modal is proposed to verify the effect of the method model.the experimental results show that Rank-1:60.9%,mAP: 49.2%,both exceeding the existing representative re-ID model.Furthermore,the adaptability of the method model to other factors in open space such as pedestrian occlusion,shooting angle,pedestrian posture,shooting light and so on is verified on MARS,the largest re-ID data set at present.The experimental results show that Rank-1: 79.5%,Map: 65.2%,which is close to the existing representative re-ID model.Therefore,the proposed method model can better extract pedestrian motion features from open-world,and has better robustness and adaptability to complex factors and variables in open-world.(2)Aiming at the problems of large amounts of parameters and computation in existing networks,a compression model without precision loss is proposed by using combination optimization strategy.With representative network frameworks such as ResNet50 and MobileNeteV2 as experimental objects,the effectiveness and universality of the combination optimization strategy are verified on re-ID standard data sets Market1501 and DukeMTMC,and finally used in the method model.the performance comparison before and after use shows that Rank-1 is improved by 0.9%,and mAP is improved by 1.9%.Furthermore,the method model is deployed on the edge device Jetson Xavier NX,and the inference results show that the energy consumption and storage space of the optimized method model are reduced by 0.9 W/h and 0.06 GB respectively,and the speed is increased by about 40%.It can be seen that the proposed combination optimization strategy is efficient and universal,and can quickly optimize and compress any network framework.The method proposed in this paper is enlightening,and the future research on re-ID can fuse multimodal information and further improve the efficiency of re-ID in open-world.