Research On Video-based Person Re-identification Algorithm

In recent years,with growing awareness of public security awareness and development of video monitoring technology,person re-identification has been one of the important research directions in the field of computer vision.Firstly,this paper introduces two fundamental methods to solve the problem of person re-identification,which are the selection of feature representation and the design of metric learning algorithm.Algorithms based on LOMO and XQDA,which is widely used in the field of image-based person re-id,is directly applied in video-based person re-identification.Although it can achieve good results,but still faces many problems.In view of above problems,this paper focuses on video-based person re-identification,and proposes two kinds of algorithms based on tempral alignment and spatio-temporal alignment.Temporally aligned pooling representation for video-based person re-identification temporally align the sequences before feature extraction and metric learning.Firstly,by segmenting and tracking the superpixels of the lowest portions of human,the motion information can be extract from a sequence.Then the motion is fitted with the sinusoid according to the intrinsic periodicity property of walking persons to select the “best” walking cycle,which can reduce the impact of noise on motion trajectory,as well as the utilization of redundant information.By means of temporally aligned pooling operation on the select cycle,not only the inconvenience to metric learning that caused by the variable frame numbers of different video sequen can be eliminated,but also the problem of increasing inter-class difference caused by variations of pose can be improved.The above method is still faced with some problems such as the inaccurate spatio alignment of features and the influences of the background.Person re-identification with spatio alignment overcomes the mentioned defects.Pedestrian images can be devided into three kind of semantic region by a deep decompositional network.When obtaining feature representation from each frame from the selected cycle,instead of the whole image,features are extracted separately for each body part,and then connect them together to represent each frame.Spatio alignment eliminates the effects of cluster background on person re-identification,at the same time,features from each body part are spatially aligned,so that the accuracy of the feature extraction and dentification process can be significantly improved.Extensive experimental results on the public datasets show that by introducing time alignment and spatio alignment to video-based person re-identification,the algorithm presented by this paper has been significantly improved in recognition accuracy compared with the state-of-the-art approaches.