Research On Person Re-identification Based On Deep Learning In Complex Environment

Pedestrian recognition is a technique that locates pedestrians to obtain their trajectory to continuously track the target pedestrian,while pedestrian re-identification is a technique that solves the problem of target pedestrians after losing their field of view in cross-camera tracking.Pedestrian re-identification has attracted the attention of many researchers in the fields of criminal investigation,intelligent security,and artificial intelligence,and has achieved rich research results.The current pedestrian re-identification method performs pedestrian identification in a single scenario,and the recognition accuracy obtained on the public data set exceeds that of human identification.However,it is difficult to confirm the identity of pedestrians in complex environments such as large differences in camera shooting angles,light changes,blurred pedestrian movements and occlusion,making pedestrian re-identification still challenging.Aiming at the above complex environment problems,this paper uses deep learning methods to study pedestrian re-identification technology from the following two aspects:1.Aiming at the problem that the accuracy of pedestrian re-recognition is low due to the large difference in camera angle,light change,and blurred pedestrian movement,a pedestrian re-recognition method based on local feature fusion is proposed from the perspective of joint learning of multiple features.In Res Net-50 networks,local learning branches and batch feature erasure branches are introduced;The local learning branch uses the method of horizontally dividing into 6 blocks to learn the characteristics of different parts of pedestrians;The Batch Feature Erasure branch allows the network to extract more discriminating local features by randomly erasing image features per batch;And use the global branch to supervise the training of these two branches.The improved network is more able to learn the characteristics of refinement and diversity.Finally,experiments were conducted on the Market-1501 and Duke MTMC-re ID datasets,rank-1 and m AP indicators are higher than most mainstream models,of which Rank-1 and m AP reached 95.06% and 84.12% on the Market-1501 dataset,respectively,which proved that the method can still identify travelers in environments such as large differences in camera angles,light changes,and blurred pedestrian movements.2.Aiming at the problem of pedestrian occlusion affecting pedestrian identification,from the perspective of paying attention to the saliency area of pedestrians,a network based on mixed domain attention mechanism is proposed.Under occlusion conditions,some of the pedestrian features are covered by occlusion features,this paper will design the CA-SA model embedded in the network based on local feature fusion,enhance the network’s learning of the relevant features of the spatial domain and the channel domain,and pass the shallow salient features into the back three branches to weaken the influence of the occluded useless features on the overall features,improve the influence of background information,and further batch erase the features of the later BEF branches to strengthen the feature learning of the unobstructed part of the network.Finally,experiments were conducted on the Market-1501 and Duke MTMC-re ID datasets,where rank-1 and m AP metrics were higher than most current methods of reiding with occluded pedestrians.Among them,the Rank-1 and m AP data sets reached 90.20% and 80.13% respectively on the Duke MTMC-re ID dataset,which proved that the CA-SA model can improve the recognition ability of pedestrian occlusion.