Research On Customized Gesture Recognition Method Based On Binocular Vision

Gestures are indispensable part in interpersonal communication, and gesture recognition also has become an important technical in human-computer interaction. In many applications of virtual reality, mouse, keyboard and other traditional means of human-computer interaction are difficult to meet the user’s interest when a user manipulate the virtual objects. Gesture recognition has broad application prospects in many applications, especially the vision based gesture recognition, it can provide a very harmonious, natural human-computer interaction. Therefore, the vision based gesture recognition has become a hot research topic. Gestures have ambiguity, diversity, complexity and other characteristics, and vision based gesture recognition has been studied in computer vision, which is more active and challenging. This thesis designs a customized gesture recognition system based on binocular vision, which is able to recognize the static and dynamic gestures.This thesis analyzes and summarizes the research of gesture definition, segmentation and recognition at home and abroad. Has an systematic research of gesture recognition models and algorithms, which are based on binocular vision. The research content and the main research results are as follows:This thesis studies imaging model of camera and some coordinate systems which camera calibration uses. Taking the camera distortion case into account, this thesis implements an adaptive camera calibration algorithm combined with the classic Zhang Zhengyou camera calibration algorithm and uses it to calibrate the binocular camera.In the gesture segmentation section, gesture segmentation use skin color model in this thesis. Because skin color has good cohesion in a specific color space. The gesture image is converted from RGB color space to HSV color space. In this thesis, the gesture segmentation also uses the depth information of binocular camera and the geometric characteristics as segmentation’s constraint conditions, which can reduce influence of similar skin color in background and the other skin except hands. This thesis improves the traditional Canny edge detection algorithm and proposes an adaptive Canny edge detection algorithm in the section of gesture edge detection.In the section of gesture tracking, the Cam-shift algorithm is used. The segmented target can use the color probability map to determine the target’s centroid.For gesture recognition, this thesis divides gestures into two types:static and dynamic gestures. Fisher linear discriminant algorithm is used for static gesture recognition. Taking the existence of the problem of occlusion into account, dynamic gesture recognition uses the binocular camera’s advantage to establish HMM. In the process of identification, it also uses two synchronous video for recognition, which can improve the recognition rate.