Research And Design Of Person Re-Identification Method For Surveillance Scenarios

With the rapid development of economy,people have higher and higher demands for safety of property and body.Thus,survillance system has drawn many researchers' attentions.With the poluparization of surveillance cameras,the data from surveillance cameras becomes enormous gradually.The traditional way of processing surveillance data by manpower is impracticable.So,intelligent video surveillance that focus on how to combine computer vision and surveillance system becomes a hot topic rapidly.Among this,person re-identification draws more and more people's attentions due to its wide range of application scenarios and research significance.Recently,with the appearance of many large-scale datasets and deep learning systems that could efficiently use them,object detection and tracking has made significant progress.However,in a scenario that people wants to find a specific target person from multiple cameras' view,the performance is still limited.Most researchers still consider object detection and person re-identification as two independent tasks in computer vision,the study of how to combine them is few.Nowadays,the task of searching a target person from cross-camera view is normally divided into two independent tasks,i.e,person detection and person re-identification.This thesis introduces an end-to-end method that uses attention model and Recurrent neural network to identify the person of interest.This thesis considers this problem as sequence modeling and let the detection process guide the identification process and combines RNN with attention mechanism.We first use ResNet to extract features of the target person,then add them to a ConvLSTM to help it choose the next focus region,combines all the observation at last to infer person who are most likely to be the target person.This method could reduce the number of images from surveillance video that need to be processed for person re-identification and performs the identification task with less computation cost and time.This thesis also compares different combinations of detection methods and re-identification methods.The method proposed above is proved to be valid by testing on several datasets.