A fast and robust head pose estimation system based on depth data

This paper proposes a performance enhancement algorithm for Kinect depth data based head pose estimation method that uses discriminative random regression forest (DRRF). In the testing phase of DRRF, patches are extracted from the whole query depth image and then are passed through each of the tree in the trained forest for head detection and head pose estimation. In this procedure, however, errors in head detection may occur when some complex background information appears in the depth image. Moreover, the more background information the depth image contains, the more processing time is required. Another drawback of DRRF is that it is very sensitive in live mode. For example, the measurement of head pose may vibrate heavily even the pose of the head stays unchanged. In this paper, we present an improved algorithm by combining DRRF with Kalman filter. The new algorithm has greatly improved the reliability for head pose estimation. In this approach, the head location is first predicted using Kalman filter, and then patches are extracted from the head region defined by the predicted head location. The head pose is then estimated by passing these patches through DRRF for regression. Finally, the noisy regression result is refined by the correcting model of Kalman filter. Experimental results show that the proposed algorithm is faster, more robust and more accurate than the original DRRF.