英文摘要 | With the development of computer sciences, the traditional Human Computer Interface (HCI) has become limited for the more and more complicated requirements. Hence, the natural and intelligent interactive means, such as voice, vision and advanced sensor based methods, have emerged as the time require. The computer vision based method is getting more and more attention for its naturalness and convenience. Moreover, hand gesture is one of the most interactive means that has been widely applied in areas such as Virtual Reality, Human Computer Interaction, Sign Language Recognition and Tele –robotics. The research on hand detection, tracking and gesture recognition in real scene has both significant academic importance and broad applications. With the aim of designing a robust Human-Computer interaction system, we focus our research on hand detection, tracking and gesture recognition, the main contributions are summarized as follows: The Adaboost statistical method, which was applied widely in face detection, was used to palm detection. Each palm has three long clear lines, which can be used as feature for detection. The advantage of this method is high speed and high accuracy. Through integral image to calculate simple feature and the cascade structure of the classifier ensure that the method work in real time. By modeling skin color space in normalized RGB space based on Gaussian model and utilizing an adaptive threshold method to remove the influence of the changing light and background, we can get the perfect skin color extraction effect in ninety percent condition. We proposed the method combining particle filter and Active Shape Models together to track the moving hand. This method units the advantage of each method and overcome their disadvantages. Particle filter, a recursive Bayesian filter based on particle set propagation, can maintain multi peaks distribution at the same time, thus can be used to the estimation of nonlinear and non-Gaussian problem effectively. Active Shape Models can represent every shape of the hand accurately. In this paper, we use particle filter to track the whole hand’s moving first, and then based on Active Shape Models to adjust each finger’s shape. This method can track each kind of the hand motion,which includes not only the whole hand’s translation, rotation and scale, but also each finger’s flexion and rotation. We implement dynamic gesture recognition utilizing the result of the tracking part. We make improvements to the existed Hausdorff distance and make it suitable for the similarity measurement in noisy environment. At last, we realize the recognition of simple static hand gesture “scissors, rock and paper” in real scene based on the modified Huasdorff distance. The result is encouraging. |
修改评论