Person Re-Identification for Mobile Robot using Online Transfer Learning

Abstract

Person re-identification (re-ID) is generally a problem of identifying the same person using different person images captured by plural cameras that do not share a visual field. The main applications of person re-ID are systems for security, marketing and so forth in commercial facilities using monitoring cameras installed outside. While on the other hand, person re-ID for mobile robots can be defined as a problem of identifying the same person using different person images at different places captured by cameras mounted on robots. This paper proposes a person re-ID method for mobile robots that periodically provide services to specific groups. This method consists of the following two components: (1) a feature extractor that learns person feature representation based on Triplet Loss from person region detected by region-based CNN, (2) a person re-identifier that learns to identify persons using person images collected while a robot moves around a room. The person re-identifier incorporates adaptive transfer learning to periodically relearn the same persons with different appearance, e.g., clothes, directions, and postures. Performance of the proposed method is evaluated by experiments using a large open dataset and a self-made dataset periodically collected for the same group by a mobile robot.