Research On Re-identification Of Occluded Person Based On Deep Learning

Person re-identification technology refers to the technology of using computer vision technology to detect whether the person under test has a unified identity under different time and space conditions.With the development of social urbanization and the establishment of a large number of basic public facilities and venues,public safety has become a topic of increasing concern to the society.Under the background of the country’s high emphasis on and pushing the development of science and technology in the field of public safety,video surveillance algorithm s are used in various large cities and transportation fields have been widely popularized and applied,among which person reidentification technology is the core technology of video surveillance algorithm.In the complex environment of public places,the difficulty and resolution of face capture become the bottleneck of identity verification and video surveillance.Person reidentification technology uses the characteristics of the whole person instead of a single face to realize person identification.However,in most cases,surveillance cameras cannot capture complete person images,and most of them have a certain degree of occlusion,which improves the accuracy of person re-identification to a certain extent.Thesis studies the key technology of person re-identification under occlusion,designs a set of real-time end-to-end person re-identification algorithm,and considers the application scene of person re-identification at night,explores the application of the algorithm in the infrared scene sex.The main content of thesis is divided into four parts,as follows:(1)Explored the existing general detection technology in the field of deep learning,and designed an algorithm suitable for person detection through the analysis of person characteristics in normal lighting environment and night infrared environment,so as to realize the precise positioning of person objects.(2)The object tracking algorithm commonly used in the field of deep learning is studied,and it is optimized for person objects and night infrared scenes,and the person detection technology is organically combined with the person object tracking technology to realize automatic identification and long-term tracking of person objects.(3)The difference between the visual transformer and the convolutional neural network is studied,and the visual Transformer and the convolutional neural network are combined to more effectively extract the features of persons under occlusion,and by designing the multi-granularity structure of the local branch and the global branch and The cross-branch attention mechanism between them realizes that different branches have different importance under different occlusion degrees,and realizes features that are more robust to occluded persons.(4)Combine person detection with person detection and identity re-identification technology to realize a high-performance real-time end-to-end person re-identification algorithm,and apply it to infrared scenes through incremental learning.The final experimental results show that the multi-grained branch structure and cross-branch attention mechanism designed in thesis can well adapt to the feature extraction of persons under different occlusion situations.In addition,thesis combines detection,tracking and recognition to launch a real-time The end-to-end person reidentification algorithm can be well applied in infrared scenes and has application value.