Real-Time Action Recognition System from a First-Person View Video Stream

摘要

This paper presents an approach to recognizing people's actions directed towards a robot, from a video collected by a camera positioned on the robot. Our approach has two modes: "stand by" mode -when a system waits for an action to start and "active" mode -when the system recognizes actions from the video stream. "Stand by" mode allows to save resources during an absence of actions and also gives a starting point of an action, which helps for recognition. Our action recognition model is based on a Bag-of-Words model, but we utilized sparse features in order to process video frames faster. As a result, the system implemented with this approach could continuously recognize actions in real-time using fewer computations. The performance of our method was experimentally evaluated and the results are given at the end of the paper.