Multi-stage adaptive regression for online activity recognition

Online activity recognition which aims to detect and recognize activity instantly from a continuous video stream is a key technology in human-robot interaction. However, the partial activity observation problem, mainly due to the incomplete sequence acquisition, makes it greatly challenging. This paper proposes a novel approach, named Multi-stage Adaptive Regression (MAR), for online activity recognition with the main focus on addressing the partial observation problem. Specifically, the MAR framework delicately assembles overlapped activity observations to improve its robustness against arbitrary activity segments. Then multiple score functions corresponding to each specific performance stage are collaboratively learned via a adaptive label strategy to enhance its power of discriminating similar partial activities. Moreover, the Online Human Interaction (OHI) database is constructed to evaluate the online activity recognition in human interaction scenarios. Extensive experimental evaluations on the Multi-Modal Action Detection (MAD) database and the OHI database show that the MAR method achieves an outstanding performance over the state-of-the-art approaches.