Video Action Recognition Using Depth Motion Maps

This thesis presents an algorithm for modifying Depth Motion Map (DMM) pyramid analysis based action recognition. It extends the work presented by Chengwu Liang to better handle incoming data. Morphological image transforms are used to remove incoming noise created by the inaccuracies in a Microsoft Kinect when the subject is stationary. This avoids the problem of motion being registered in frames where there is no motion. It also uses a non-linear cost metric on the motion energy to ensure some degrees of shift invariance in the Z axis, which is towards or away from the camera. In addition, advanced information fusion methods are adopted to improve feature representation, leading to better performance. Experiments are conducted on video data containing a variable number of stationary frames at the start and end of an action. The experiments are conducted with many different classifiers. The results show that a performance benefit is realized due to the use of these enhancement methods.