Research On Pedestrian Re-identification Based On Deep Learning

Aiming at insufficient occlusion samples and strengthening the dicing area association,an enhanced PCB network is proposed in this thesis for pedestrian recognition.First,regarding insufficient occlusion samples,the traditional image enhancement combined with random erasing is used to pre-process the image,which can make up for the shortcomings of traditional image preprocessing methods and improve the generalization ability of the model.Secondly,for the hard partitioning method of PCB network,by improving the network structure and increasing the connection between adjacent dicing blocks,the accuracy of the model can be improved with the RPP network.Based on two improvements,a series of experiments were carried out on the public dataset Market1501,DukeMTMC-reID,CUHK03.The experimental results show that the proposed two enhanced methods can effectively improve the person re-identification performance.Under the same test conditions,the performance of the PCB and RPP algorithm was improved by 4.18%,0.61%,and 1.27%,respectively.Aiming at the problem that time attention mechanism ignores useful imformation in low quality images,an improved time attention mechanism model is proposed.The spatial self-attention mechanism is added to the image feature extraction,which can effectively pay attention to the internal spatial information of the image.At the same time,the random erasing algorithm is adopted in video frame image preprocessing.The experiment on the public data set iLIDS-VID combined with the two methods shows that the person re-identification rate in complex background can be improved,and Rank-1 can reach 74.4%.In this paper,the video frame image after random erasing preprocessing is taken as input,and the performance of RNN variant in feature fusion is compared based on CNN and RNN model.A series of experiments were carried out on the public dataset iLIDS-VID.The experimental results show that the bidirectional recurrent neural network combined with the random erasing algorithm can well fuse the features of the complex background video sequences and effectively solve the image occlusion problem.