Video-based Person Re-Identification

Person re-identification(re-ID)aims to identify a given pedestrian in different cameras.Given an image or video of a specific person,the process of person re-ID is to determine whether this person has been observed by another camera.Due to the background noises,illumination variations,viewpoint variations and occlusions between different images of the same person,the accuracy of person re-ID cannot been applied in daily life.A large amount of work has focused on two aspects: one is feature extraction,which obtains a robust descriptor to extract reliable features of pedestrians;another is to learn a new metric distance space to compute the similarity between different person.Since more large-scale datasets have been introduced,CNN-based deep learning models become increasingly popular and achieve good performance in person re-ID.According to the data resources,person re-ID can be divided into image-based person re-ID and video-based person re-ID.Video-based person re-ID provides more information about pedestrians,but how to aggregate useful information of all frames is still an open issue.In existing literature,person re-ID with single image has been explored extensively.This paper mainly focus on the research in video-based person re-ID.To sum up,main contributions of this paper are summarized as follows:(1)Through the learning of basic methods in image-based re-ID and video-based re-ID,we conclude the merits and drawbacks of all kinds of related methods and figure out how to improve the performance in videobased re-ID.(2)According to the feature fusion methods,we combine the framelevel features and video-level features to form a more robust descriptor.After a relative simple feature extraction and metric learning,we could select the possible candidates for further process.We apply Flow Energy Profile(FEP)signal to select more discriminative fragmentations from the whole video sequences and then extract 3DHOG features.Compare the fusion features using top-push distance model,we achieve the rank-1 accuracy of 58.4% and 60.0% on datasets iLIDS–VID and PRID2011 separately.(3)Although person re-ID based on deep learned network can learn comparable robust features automatically,they rely more on global descriptors and ignore similar details between different identities.This paper designs a multi-task model combining the aggregation method through region-based quality prediction and attributes recognition,which can reduce the impact caused by occlusion and misalignment.