Comparison of shape descriptors for hand posture recognition in video

Hand posture recognition remains a challenging task for in-line systems working directly in the video stream. In this work, we compare several shape descriptors, with the objective of finding a good compromise between accuracy of recognition and computation load for a real-time application. Experiments are run on two families of contour-based Fourier descriptors and two sets of region-based moments, all of them are invariant to translation, rotation and scale changes of hands. These methods are independent of the camera view point. Systematic tests are performed on the Triesch benchmark database and on our own large database, which includes more realistic conditions. Temporal filtering and a method for unknown posture detection are considered to improve posture recognition results in case of video stream processing.