Human action recognition using histograms of oriented optical flows from depth

Official URL: http://dx.doi.org/10.1007/978-3-319-14249-4_60

In this paper we develop a new method for recognizing human actions from depth data. 2D optical flows from depth images are computed for the entire action instance. From the resulting optical flow vectors, patches are defined around each joint location to learn local motion variations. These patches are grouped in terms of their joints and used to extract a new feature called ‘Histograms of Oriented Optical Flows from Depth (HOOFD)’. In order to encode temporal variations, these features are generated in a pyramidal fashion. At each level of the pyramid, action instance is partitioned equally into two parts and each part is employed separately to compute histograms. Oriented optical flow histograms are utilized due to their invariance to scale and direction of motion. We performed several experiments on publicly available databases and compared our approach with some of the state-of-the-art methods. Results show the success of the proposed method.