Person Re-Identification from different views based on dynamic linear combination of distances

Abstract

Person re-identification from videos taken by multiple cameras from different views is a very challenging problem that has attracted growing interest in last years. In fact, the same person from significant cross-view has different appearances from clothes change, illumination, and cluttered background. To deal with this issue, we use the skeleton information since it is not affected by appearance and pose variations. The skeleton as an input is projected on the Grassmann manifold in order to model the human motion as a trajectory. Then, we calculate the distance on the Grassmann manifold, in order to guarantee invariance against rotation, as well as local distances allowing to discriminate anthropometric for each person. The two distances are thereafter combined while defining dynamically the optimal combination weight for each person. Indeed, a machine learning process learns to predict the best weight for each person according to the rank metric of its re-identification results. Experimental results, using challenging 3D (IAS-Lab RGBD-ID and BIWI-Lab RGBD-ID) and 2D (Prid-2011 and i-LIDS-VID) benchmarks, show that the proposed method can boost re-identification ranking thanks to its ability to define the optimal weight for each person independently of view and pose changes.