Studies On Appearance-Based Hand Gesture Modeling Method Under Complex Background

With the emergence of new hardware and development of new application fields, human-machine interactive activities has become more and more important in people's daily activities. The traditional interactive ways, including keyboard, mouse and light pen, limit the speed and naturalness of man-machine communication. While, the interactive ways based on computer vision can solve above problems better, and are more in conformity with the way of people thinking. However, human hands possess diversities, multi-meanings and differences in time and space. and people hands themselves are complex variable forms, so the study on gesture modeling methods based on computer vision is a challenging interdisciplinary research subject.A complete gesture model system based on computer vision includes gesture video acquisition, gesture segmentation, and gesture modeling. Gesture modeling consists of gesture analysis and gesture recognition. Through summarizing and analyzing various methods in gestures modeling process,it must be pointed out that accurate gesture segmentation and robust gesture feature extraction are keys of gestures modeling. The concrete research works in this paper are as following:First , this paper gets gesture streaming video through camera, separates streaming video into frames without redundant, extracts gestures image frames from streaming video, and at last wins a succession of gestures image sequence.In gesture segmentation, aiming at the limitation of single clue segmentation algorithm, segmentation method combinating color information and sport information is adopted. Firstly it uses elliptic color model based on KL transform for skin detection and improves the precision of skin detection. Then, three-frame difference method is used for continuous gesture image sequence, and the method improves the blocking problem of second-frame difference method. At last, two results are combined and gestures segmentation results are ideal.As to feature extraction phase, for the whole matching cannot deal with partial sheltering and lack of specificity, a method of Polar-Radius-Invariant-Moment based on key-point is proposed to extract features for shape recognition. It extracts key-points of gesture shape outline through discrete curve evolution (DCE) method, and uses Polar-Radius-Invariant-Moment based on key-point for describing gesture shape features. Experimental results show that the new method has very good classification properties, and can maintain good robustness even if the object contour is ill-segmented or noisy.