Motion detection and action recognition using elastic registration

Motion detection has been widely used as a fundamental technique in automated video surveillance, intelligent transportation, real-time video monitoring and secure communication systems. In addition to object movement, action recognition (gesture/ posture etc) play a vital role to discriminate the type of activity in applications related to sports (Cricket, Football, Archery etc.), health (behavior analysis in patients), and law enforcement agencies (suspect behavior analysis) etc. In this research, we propose to detect motion (shift in position) using modified normalized phase correlation based on non-rigid registration of two successive images/frames of video to detect any elastic movement in the particular scene. Moreover, gram-polynomial based image decimation is used to reduce the computational complexity of the proposed method. After detection of motion, object can be tracked using accumulative non-rigid translational shift provided by modified normalize phase correlation. These shifted points (pixel movements) can be converted to unit lengths based on camera parameters (angle of view, resolution etc.). However, proposed method is used to recognize actions (gesture/posture etc.) which in turn discriminate the nature of activity such as type of sport (Archery etc.). The key to distinguish the actions of each sort of sport will be pattern of accumulative translational shift in pixels provided by the proposed method. The proposed method is efficient to the extent of near real by taking 0.82 second aggregated time for both registration and complete processing of an image using 2.2GHz Core i3 processor speed and 2GB of RAM.