Multi-rate recognition of human gestures by concurrent frame rate control

In order to sufficiently represent and recognize human gestures, one needs, in the first place, to know at what frame rate a gesture should be sampled. To investigate the temporal-domain problems on gesture recognition, a framework is proposed that recognizes arbitrary gestures at multiple frame rate by running recognition processes in parallel. In our system, each recognition process runs on-line and in real-time by controlling the frame rate autonomously. As a benchmark image database, we use the ViHASi (Virtual Human Action Silhouette) database which is publicly available. We created an original image database that contains 18 kinds of sporting and stretching gestures, both are to evaluate the recognition performance under different frame rate conditions. By examining the recognition rates of each gesture at each frame rate, we found gestures that exhibit frame rate dependency and the gestures that do not. We experimentally demonstrate that the dependency is largely determined by the distinctiveness among target gestures.