Research Of Video Person Re-identification Based On Metric Learning And Transfer Learning

Video-based person re-identification is an important task in video surveillance.Among the existing video-based person re-identification methods,the metric learning based ones have achieved very interesting results by further improving the discriminability of the original video features.However,there still exist some shortcomings in these methods.For example:(1)the amounts of discriminant information contained in different negative samples are usually inconsistent,so their contributions to metric learning are also different.However,these metric learning based methods cannot deal with different negative samples effectively.(2)These metric learning based methods usually rely on a large number of labeled data in the training process.In practice,it is a time-consuming and labor-consuming work to collect and label a large number of pedestrian videos from non-overlapping cameras.In view of these problems,we have carried out two-stage researches in this paper.First,a Negative sample Sensitive Metric Learning(NSML)approach is proposed.In the process of measurement learning,different penalty factors are applied to the selected negative samples,to solve the problem that different negative samples have different contributions to metric learning.In addition,we propose a Transfer Learning based Unsupervised Metric Learning(TLUML)approach.TLUML first projects the source and the target data into the common subspace,and then learns the distance metric suitable for target data in the common subspace.In order to verify the effectiveness of the proposed method,we compare the proposed approaches with several existing video-based methods.The experimental results demonstrate that the proposed NSML can make better use of the useful information contained in difficult negative samples.The proposed TLUML can learn the matching model suitable for the target data by using the label information contained in the source dataset.