Scale-based human motion representation for action recognition

The design of action recognition algorithms often relies on knowledge of the particular problem, which is not always available. Moreover, algorithms usually incorporate a number of parameters, which influence their performance. To solve these problems, the possibility of developing more general action recognition algorithms by systematic reduction of complexity of human motion, instead of designing more and more complex algorithms is explored. The key for reducing complexity in systematic decomposition of human motion to different scales, each representing different level of motion detail is seen. A human action is assumed to influence different scales of motion, and they should be observed that way. The features, obtained on different scales of motion can be joined together to represent the complex motion in uniform and manageable way. Our approach was tested in the sports domain, on a particular problem of detecting the action of athlete hitting the ball with a racquet in the game of squash. Video recordings of actual tournament match were used, and manual annotations were provided by squash expert as a ground truth.