Human Action Recognition Using Optical Flow Accumulated Local Histograms

This paper addresses the human action recognition task from optical flow. We develop a non-parametric motion model using only the image region surrounding the actor making the action. For every two consecutive frames, a local motion descriptor is calculated from the optical flow orientation histograms collected from overlapping regions inside the bounding box of the actor. An action descriptor is built by weighting and aggregating the estimated histograms along the temporal axis. We obtain a promising trade-off between complexity and performance compared with state-of-the-art approaches. Experimental results show that the proposed method equals or improves on the performance of state-of-the-art approaches using these databases.