Research On Video-based Person Re-identification Method By Multi-level Deep Feature Representation

Person re-identification is a study of identifying and matching the same person across non-overlapping camera views.Person re-identification technology is mainly used in video surveillance and criminal investigation.It has important academic significance and application value in the field of computer vision and public safety.Due to the rapid development of current video big data and the urgent need of intelligent video analysis technology,this thesis focusing on video-based person re-identification technology by combining multi-level deep representation learning.Two methods are proposed innovatively from the perspective of the staged model and the end-to-end learning model to effectively improve the accuracy of person re-identification.From the perspective of the stage model based on deep feature representation and distance metric,this thesis proposes a joint person's multi-level deep feature representation network and ordered weighted distance measure fusion algorithm for video-based person re-identification,which is aimed at solving the problem of person appearance and behavior similarity,and mismatch caused by the distance difference between same persons under different features.In the stage of person feature representation,the proposed person multi-level deep feature representation network can not only learn the space-temporal features of person in the video sequences,but also acquire the global appearance features and local appearance features of person.In the stage of ordered weighted distance fusion algorithm,this thesis inputs the feature representations of person into the distance metric learning,and calculates the independent distances of person under the three types of features.Then,the algorithm sorts the distances,and optimizes the distance weights according to the result of distance ranking.Finally,this method combines the three types of distances to get the final distance,and to match the person accurately.Extensive experiments comparing with state-of-the-art methods on two public datasets show that the proposed algorithm not only improves the recognition rate of video-based person re-identification,but also increases the abundant and complete feature representation ability of person.From the perspective of end-to-end model based on hybrid deep features and two-stream Siamese networks,this thesis proposes a video-based person re-identification method called as the end-to-end learning architecture with hybrid deep appearance-temporal feature,which can focus on how to establish the stable person appearance feature representation model that can eliminate the influence of person's interference frames in video and how to effectively use the complementary features of appearance features and temporal features in video to calculate the similarity between persons.The end-to-end deep learning architecture consists of the hybrid deep feature structure and the two-stream Siamese network,which can learn the appearance feature of person's pivotal frames,the temporal feature of video sequences,and the independent distance metric similarity of different features.Extensive experimental results on three public datasets show that the proposed architecture not only improves the feature representation ability of person,but also makes the recognition rate of the proposed algorithm better than the state-of-the-art algorithms.