| With the development of video surveillance systems,the video data from surveillance sys-tems accumulate more and more abundant.It is significant to build an efficient and intelligent video security surveillance system for public security management.Person re -identification technology aims at matching person images with the same identity in non overlapping cameras,which is a critical technology for building an intelligent security monitoring system.In the video person re -identification task,it is critical to extract discriminative features from person video data and design the generalized models in real-world surveillance systems.Aiming to extract discriminative features from person video sequences,this thesis proposes a method based on adaptive normalization.By adding adaptive normalization modules to the deep network model,the degree of feature dispersion is improved in the normalization process,which ensures that it strives to boost the representative ability of the extracted features.Aiming to improve the generalization ability of the network model,this thesis proposes a domain augmentation normalization method.By introducing the mechanism of adaptive dy-namic adjustment for the different feature distributions,the diversity of the feature distribution is augmented.Besides,the network model based on domain enhancement and normalization is applied to boost the generalization ability of the network model in unknown domains.Extensive comparison and ablation experiments were conducted on four well -known pub-lic datasets,and the results of experiments prove the rationality and effectiveness of each inno-vation.Specifically,the domain augmentation normalization method achieves state-of-the-art in the video-based person Re ID domain generalization tasks. |
