Action Recognition by Shape Matching to Key Frames

Human action is, in general, characterized by a sequence of specific body postures. An action is often recognizable from a single view of an action specific posture. In this paper we demonstrate that specific actions can be recognized in long video sequence by matching shape information extracted from individual frames to stored prototypes representing key frames of the action. The matching algorithm is tolerant to substantial deformation between image and prototype and recognizes qualitatively similar image shapes produced by the body postures. The algorithm is applied to the recognition of specific tennis strokes in up to 2 min long video sequences. Using key-frames within the sequence all (1015) forehand strokes are found with no false positives. The backhand strokes are detected with somewhat lower performance. We conclude that this type of approach to action recognition has large potential for application to automatic video analysis and editing.