Improved Metric Learning Algorithm for Person Re-Identification Based on Asymmetric Metric

Person re-identification(re-ID) is becoming a hot research topic because of its value in both machine learning and video surveillance applications. In order to improve the robustness of Metric Learning by Accelerated Proximal Gradient(MLAPG), a person re-ID algorithm, called Asymmetric -MLAPG, is proposed on the basis of asymmetric metric. Unlike traditional metric learning which ignores the taking environment of the person images, asymmetric metric learning aims to find the feature matrix of each camera and map the vectors of images to a common space. In this paper, we rebuilt MLAPG model by Asymmetric metric. In addition to adding a regularization term which controls the influence of inconsistency of metric, there are two more regularization terms in order to control the offset of feature matrix from the initial value. Experiments show that the Asymmetric -MLAPG algorithm can achieve the better recognition rate and wide applicability on commonly used person re-ID data sets.