Research On Intra-Video Variation For Video-Based Person Re-Identification

In order to protect public safety,surveillance equipment is increasingly popular in public places.Person re-identification is derived from video surveillance,which is used to determine whether pedestrians captured in non-overlapping perspectives are the same target.In recent years,it has become a hot topic in the field of computer vision.This paper mainly studies to effectively extract and process the information of pedestrians in the video sequence.To solve the problems of multi-camera surveillance such as ilumination changes,viewpoint changes,we use convolutional neural network(CNN)and recurrent neural network(RNN)to extract information.In order to measure the similarity between the two pedestrians,we use the Siamese Network to obtain the features of two pedestrians,and evaluate the similarity through the Siamese Loss.The Siamese Loss is mainly used to solve large variations between spatial-temporal features extracted from different videos(called inter-video variations).On this basis,we introduce the concept of 'mean-body' to solve large variations between different spatial-temporal features extracted from the same video(called intra-video variations).Specifically,mean-body herein represents the center of gravity of the samples of human bodies within one video sequence.It seeks to capture the variation of a person in the whole video sequence.The intra-video constraint emplbys the distance between each frame and the mean-body to minimize the intra-video distance within a video sequence.We combine the mean-body,intra-video loss and Siamese Network to form a new network,so that the network can reduce inter-video variations and intra-video variations simultaneously.The main contributions of this paper can be summarized as follows.1)We introduce the concept of'mean-body' to represent individual human bodies in a holistic manner.2)An intra-video loss is defined to address the intra-video variation for person re-identification.3)We combine the intra-video loss and the Siamese loss to train the whole network,which proved to be superior over existing work in terms of re-id accuracy.