Action Recognition in Semi-synthetic Images using Motion Primitives

This technical report describes an action recognition approach based on motion primitives. A few characteristic time instances are found in a sequence containing an action and the action is classified from these instances. The characteristic instances are defined solely on the human motion, hence motion primitives. The motion primitives are extracted by double difference images and represented by four features. In each frame the primitive, if any, that best explains the observed data is identified. This leads to a discrete recognition problem since a video sequence will be converted into a string containing a sequence of symbols, each representing a primitive. After pruning the string a probabilistic Edit Distance classifier is applied to identify which action best describes the pruned string. The method is evaluated on five one-arm gestures. A test is performed with semi-synthetic input data achieving a recognition rate of 96.5%.