Research On Person Re-identification With Part Based Model,Group Affinity And Data Augmentation

Video surveillance plays an important role in security alerts,tracking suspects,and finding missing people.And person re-identification targets at modelling person identity information in a large scale video surveillance system,which is of great significance for tasks such as target tracking and behavior analysis.However,The main difficulty for the Re-id task resides in the large variations of visual perspectives,illuminations and poses,which may lead to difference in the appearance of an individual.To solve above problems,we focus on three aspects:extracting pedestrian feature,utlizing group affinity information and expanding data set and effectively overcome the impact of difference in the appearance of an individual.The main contribution of this paper are summarized as follows:1.A new aligned part model for person re-identification is designed,which inherits the high-efficiency expression of the part model,and introduces the spatial transformation network structure to align the feature map of pedestrian.The proposed method can alleviate the offset of the feature map caused by the occlusion,inaccurate position of the component detection frame,and further improve the discrimination and robustness of features.2.A powerful deep learning and probability hypergraph based model is proposed,which aims at fully utilizing group affinity information for more discriminative features.Specifically,hypergraph model for group information mining is novelly proposed in the training phase,which significantly improves the performance of features acquired by the baseline networks.Extensive experiments on two popular benchmarks also demonstrate applicability on different baseline networks.3.A data set expansion method for person re-identification is designed,which generates pictures by deep generative adversarial networks,which can help to alleviate the overfitting problem caused by the insufficient data size and the similar style of the data set.At the same time,the label smooth normalization method is used to determine the label of the generated image,so that the generated and the real images can be sent to the neural network for training,thus the robustness and the discriminative of the features can be improved.