Hough-based action detection with time-warped voting

Hough-based action detection methods cast weighted votes for action classes and positions based on the local spatio-temporal features of the given video sequences. Conventional Hough-based methods perform poorly for actions with temporal variations because such variations change the temporal relation between the local feature positions and the global action positions. Some votes may scatter because of such variations. In this paper, we propose a method for concentrating scattered votes through a time warping of the votes. The proposed method calculates the offsets between the scattered voting positions and the concentrated positions based on the votes generated through the conventional Hough-based method. The offsets warp the scattered votes to concentrate them, and provide a method of robustness even in the presence of temporal variations. We experimentally confirmed that the proposed method improves the average precision for the UT-Interaction dataset compared with a conventional method.