Distance metric learning for multi-camera people matching

In this paper, we propose a supervised distance metric learning method for the problem of matching people in different but non-overlapping camera pictures, which is an important and challenging problem for behavior understanding. Different from previous methods, which try to extract good visual features, in this paper, we try to model it as a distance metric learning problem. We formulate the problem so that the learned distance between the a pair of true matched people´ image is smaller than that of a wrong matched pair.We conducted experiments on one benchmarking dataset, and demonstrate the advantage of the proposed distance learning models over state-of-the-art multi-camera people matching techniques.