Research On Person Re-identification Based On Distance Weighted Sampling And Dynamic Image

Person Re-identification(Person Re-ID)is a research hotspot in the field of intelligent video analysis.As the scale of data continues to increase,the research on person re-identification based on traditional machine learning will enter the bottleneck area.However,with the breakthrough of deep learning algorithms in the fields of machine learning and computer vision,deep learning has became the mainstream research method for person re-identification.According to the different research objects,the research problems of person re-identification are divided into two categories:one is based on single-shot image for person re-identification,and the other is based on multi-shot video sequences for person re-identification.Person re-identification algorithm research mainly focuses on two aspects of representation learning and metric learning of pedestrian images or video.Although several important results have been obtained in these two aspects,there are still many problems,such as similarity matching between descending images of complex backgrounds or feature representation of pedestrian video sequences.In response to the above questions,Our main work includes two aspects:(1)For person re-identification of single-shot images,we proposes a distance weighted sampling algorithm and designs a triplet loss variant based on the algorithm.Specifically,an anchor image is first given and a difficult positive sample pair is selected,and then the triplet negative sample is uniformly selected according to the distance weighted sampling algorithm.The experimental results show that the proposed algorithm is superior to other pedestrian recognition algorithms based on the triplet loss function.(2)For person re-identification of multi-shot video sequences,we proposes a person re-identification algorithm based on dynamic image.The algorithm aims to represent the pedestrian sequence into the form of dynamic image by using the ranking support vector machine model,And extract the dynamic image as the input of the deep embedded learning network to extract the temporal feature of the effective pedestrian sequence.,and than combines the apparent feature of the pedestrian sequence into the final feature representation of the pedestrian sequence.Finally,the appropriate similarity measure is used to learn the matching of the pedestrian video sequence.The experimental results show that the dynamic image algorithm can effectively extract the temporal characteristics of pedestrian sequences and achieve high precision on three common pedestrian datasets.