Stereo vision-based recognition of nonwearable pointing gesture
基于立体视觉的非穿戴指势识别方法

Based on the differences between the color intensities of R, G and B components on a color image in shadow regions, we develop a novel approach to the pointing object segmentation across a clutter background. Because the wavelet transformation is with outstanding local characteristics both in temporal and spatial fields, we suggest extracting the video pointing objects based on the combination of background subtraction with the wavelet multi-scale transformation. The proposed algorithm does not require the information which is necessary for existing methods in literature, such as scene learning and training, manual calibration and a priori hypothesis. It is also robust to dynamic scene variation, shadow and noise disturbance. Based on the biological structure characteristics, we employ the position of the human head-top instead of that of the human eyes in the pointing object segmentation. This is because that the human head-top is not easily occluded by other parts of the body, and is free from the effects of the facial orientation, posture and illumination variation. This provides the flexibility and casualness for the human-computer interaction(HCI), and ensures a high speed for the interaction. For the correct matching of finger tip characteristics, we present a stereo matching strategy based on the geometric constraints among the pointing arm finger tip, the central axis of the pointing arm and the corresponding epipolar line. A reverse matching criterion is employed to ensure the validity of the processed matching. Experiment results indicate that the developed approach is efficient for the recognition of the flexible and casual nonwearable pointing in the human-computer interaction.