People Re-Identification in Service Robots

摘要

People Re-Identification (Re-ID) is essential for service robots to have effective human-robot interaction. This article proposes a Re-ID-based multi-people tracker suitable for mobile robots. It combines existing methods such as a people detector, a people localizer, a Re-ID feature extractor and a Kalman filter framework with novel data association and track management approaches. A new and publicly available RGB-D Re-ID multi-people 3D tracking dataset with ground truth is also presented. This dataset was recorded with a moving camera in an environment with obstacles, target occlusions and appearance changes. Experimental evaluation shows that the method achieves good tracking and re-identification performance on the proposed dataset at a high frame rate and outperforms a state-of-the-art method on an open-space dataset. The proposed system is computationally lightweight, robust and suitable for real-world applications, allowing for an improvement in human-robot interaction.