An HMM-Based Approach For Gesture Recognition Using Edge Features
We describe a Hidden Markov Model approach to gesture recognition employed for the ChaLearn Gesture Challenge. An HMM is constructed in which states correspond to frames of gesture templates (using the depth images only). Frames of a gesture are represented using the points in the image that are identified as lying along an edge. Each edge point is associated with several features including its X/Y coordinates, its orientation, its "sharpness", its depth and its location in an area of change. We constructed a simple Gaussian model that estimates the probability of an edge point in an observed frame corresponding to a particular edge point in a gesture template frame. The probability of one frame matching another is estimated as the joint probability of each edge point in one frame matching its nearest neighbor in the other. The most likely gesture sequence for a movie is found with a Viterbi search, as is typical in HMM systems.