Research On Depth Data Fused Human Computer Interaction Hand Gesture Recognition

Image based biometric receives more and more attentions in recent years. Human biometric is one of the most important parts in biometric, which can be used widely in natural human computer interaction, intelligent video surveillance and virtual reality. Further, the research on human hand posture and gesture recognition belongs to the research on human biometric, that segments, tracks and recognizes series of hand gestures and understands the meaning of the gestures finally. However, due to individual differences among difference hands, intricate deformation, spatio-temporal variability and the inherent ill-posedness of visual problems, hand gesture recognition becomes highly challenging.Classic hand recognition research is composed of three stages:hand segmentation, tracking and recognition. The three stages correspond to image analysis and image understanding level in computer vision. The hand segmentation stage belongs to image analysis level, and it labels the hand(s) pixels in images. This stage is the start point and the most important part, whose results will have the direct influence on the following stages. The hand tracking stage also belongs to image analysis level, and it labels hand(s) pixels in each frame in image sequence. The hand recognition stage belongs to image understanding level, and it represents the image into series image eigenvectors and classifies these eigenvectors in feature space. In this way, hand gestures will be classified and recognized.On the other hand, multi-modal data fusion field believes that single type of sensor can only acquire incomplete Information of measured object and have environmental instability, and a variety of sensors data can improve the reliability of the system. Based on these views, depth date is introduced into visual data. The hand segmentation, tracking and recognition algorithms will be studied based on visual and depth data fusion.Dynamic depth threshold based hand gestures segmentation algorithm. First of all, the skin gaussian model (mean and variance) is constructed by using MCG-Skin. Secondly, the body depth gaussian mixture model is constructed according to the depth information of the body. Thirdly, the rough dynamic depth threshold can be get, with which hand gestures can be labeled roughly. Lastly, the skin similarity of each pixel can be get by putting the rough hand gestures image into the skin Gaussian model, and Otsu’s algorithm is applied on the skin similarity image to acquire the hand gesture segmentation result. The availability of the algorithm is investigated by different perspective and experiments.Weight shift resampling based hand gestures tracking algorithm. Firstly, Relative depth histogram and its similarity measure are proposed, which is the hand gesture template for tracking. Secondly, to be aimed at particle degradation in traditional particle resampling algorithm, a new resampling method is proposed that resamples particles by non-decreasing posterior probability density gradient. This method do not delete particle in the process of resampling which can keep the diversity of the particles. The availability of the algorithm is investigated by different perspective and experiments.Relative radial distance based hand gestures recognition algorithm. To be against category problem when hypersphere support vector machines intersect the hypersphere, the relative radial distance based is proposed. This method determines the category of feature points by the ratio between the distance feature points to the center of the hypersphere and hypersphere radius. The feature points belong to the hypersphere which make the ratio minimum. The availability of the algorithm is investigated by different perspective and experiments.Finally, a hand recognition oriented image dataset is constructed. Furthermore, On the basis of the above research, a hand recognition toolbox based on Matlab is designed and implemented which can segment, track and recognize hand gestures.