Progressive Filtering Approach for Early Human Action Recognition

Human action recognition plays an important role in vision-based human-robot interaction (HRI). In many application scenarios of HRI, robot is required to recognize the human action expressions as early as possible in order to ensure a suitable response. In this paper, we proposed a novel progressive filtering approach to improve the robot's performance in identifying the ongoing human actions and thus to enhance the fluency and friendliness of HRI. Human movement data were captured by a Kinect device, and then the human actions were constituted by the refined movement data using robust regression-based refinement. Motion primitive, including both spatial and temporal information concerning the movement, was considered as an improved representation of action features. Then, the early human action recognition was accomplished based on an improved locality-sensitive hashing algorithm, by which the ongoing input action can be classified progressively. The proposed approach has been evaluated on four datasets of human actions in terms of accuracy and recall curves. The experiments showed that the proposed progressive filtering approach achieves high recognition rate, and in addition, can make the recognition decision at an earlier stage of the ongoing action.