Research On Person Re-identification In Complex Scenes

Person re-identification is one of the important research topics in the field of computer vision,which aims to study the recognition of specific pedestrians under cross-camera.Person reidentification has important research significance.It has a wide range of applications in intelligent video surveillance systems,intelligent security and transportation,and has gradually become an important means to maintain public safety and social stability,such as searching for criminal suspects,cross-border tracking and trajectory analysis.In recent years,person re-identification has made great progress,but it still faces many challenges in open video surveillance scenarios,such as view point changes,illumination changes,similar pedestrians,and occlusion,among many others,which are difficult to meet the needs of practical applications,and the accuracy needs to be further improved.In view of the above problems,this thesis adopts the method based on deep learning to study the improvement of the performance of the person re-identification model in complex scenes.The main contents are as follows:(1)Based on the backbone Res Net-50,a spatial and channel attention module(PCAM)is introduced and a multi-level feature cascade structure is constructed.PCAM focuses on improving the model’s ability to recognize different spatial location information,the multi-level feature cascade structure cascades the features of different stages to realize the supplement of detailed information,thereby improving the accuracy of the model.The results on three largescale person re-identification datasets show that introducing PCAM and building a multi-level feature cascade structure can improve the performance of the model.(2)Taking Res Net-50 as the backbone,based on the multi-level feature cascade structure,the split attention mechanism and the batch normalization bottleneck are integrated.The split attention mechanism aims to fuse multi-scale features and establish the interaction between feature channels,and the batch normalization bottleneck can solve the problem of inconsistent distribution of features in the embedding space when the model is jointly optimized with classification loss and triplet loss.The results show that embedding the split attention mechanism and batch normalization bottleneck into the model can improve the accuracy of the model.(3)Taking Res Net-50 as the backbone,based on the multi-level feature cascade structure and batch normalization bottleneck,the self-calibrated convolution is introduced and a combined pooling strategy is proposed.The self-calibrated convolutions aim to establish dependencies between feature channels and extract long-range spatial information in images,the combined pooling strategy combines average pooling and max pooling to make the model pay more attention to local salient regions in the image.Extensive experiments are performed on the three datasets,and the results show that the accuracy and average accuracy of the proposed model in complex scenes are better than most current mainstream models,and it has better robustness.(4)Based on the integrated development platform Pycharm and the programming language Python,the Pytorch deep learning framework and the cross-platform toolkit Py Qt5 are used to design the graphical interface,complete the development of the person re-identification software,and realize the person re-identification function.