Simultaneous Action Recognition and Localization Based on Multi-view Hough Voting

One problem of conventional action recognition is that it requires both human detection and human tracking before recognition. Human pose and motion vary depending on the person's action, and such variances can complicate detection and tracking. To solve this problem, previous work has proposed simultaneous action recognition and localization using Hough voting. In this paper, we present an approach that simultaneously recognizes and localizes human actions from multi-view video sequences based on Hough voting. Multi-view videos have an ability that gives robustness to the changes of human orientation and occlusion. Our proposed approach independently votes for the action labels and positions in each view and integrates them using homographical transformations. We evaluated our approach on the IXMAS dataset and confirmed that it achieved high accuracy in action recognition, localization, and robustness to the changes of human orientation and occlusion. The contribution of this paper is that it enables multi-view action recognition without advance human detection and tracking.