HSGA: A novel acceleration descriptor for human action recognition

Tracking dense features has become one of the most popular methods for human action recognition. Proper descriptors should be used to capture the motion information contained in these trajectories and motion boundary histogram (MBH), which encodes velocity information, gives best performance among state of art action recognition descriptors. In this paper, we propose to use a new descriptor, histogram of spatial gradient of acceleration (HSGA) in combination with MBH to describe actions. Our new descriptor combination is based on studies which reveal that acceleration is as important as velocity in motion description. HSGA is computed by taking histogram of orientation of spatial gradient of optical acceleration in a 3D space-time block divided into cells around dense trajectories. Optical acceleration is obtained by taking time derivative of optical flow. This combination of descriptors gave good performance on a variety of data sets. Combining these motion descriptors with a scene descriptor like HOG further improved the recognition accuracy for realistic action datasets.