Research On Intelligent Interaction Methods For A Humanoid Robot With Vision System

With the development of science and technology, robots walk into our work and lives gradually, and the humanoid robots with human features are accepted more widely and easily, which developed rapidly. As the advanced stage of robot development, Humanoid robot has become a hot area of current research because of its high degree of intelligence.The research of this dissertation focused on some methods of intelligent interaction for humanoid robot base on vision system. Firstly, a humanoid head is designed and implemented as a mechanical and control platform for humanoid robot vision system. Secondly, the architecture of vision system is proposed and some key technologies are carefully considered. Thirdly, a new method is presented to deal with the three phases of static hand gesture recognition and is applied to human-robot interaction based on hand gestures. Finally, CUDA is introduced to improved performances and real-time of vision system.The main results and novelties of the dissertation include:1. A humanoid head with 2-DOF is designed and implemented, which is controlled by servo-controller. This head has features like flexible control, accurate positioning and simple structure, but realizes general functions of humanoid head movements. With the theory of robot kinematics, the kinematical modeling and analysis of humanoid head is proposed, which can be applied to notice and trace of the region of interest.2. Vision system is the main source of information from external environment, main basis of robot's decision-making and a feedback approach to get its own behavior information. According to the deficiency of researches for vision system in domestic, an intelligent, open and integrated vision system architecture is put forward. This architecture includes image pre-processing layer, feature extraction layer, decision-making layer, and application layer. Some key technologies of vision system, such as binocular calibration and stereo vision, human face detection, color feature training and recognition, motion detection, and selective perception, are researched and implemented. A method for color feature training and recognition is presented, which possesses real-time efficiency and simple implementation. Also, a method for selective perception based on vision is proposed, which comprehensively considers effects of target size and distance for human and non-human respectively. In addition, perception recession is introduced to avoid the time of humanoid robot staring at an object is too long.3. A new method is proposed to deal with the three phases of static hand gesture recognition. Firstly, we improve the traditional RCE neural network and apply it to hand image segmentation. The improved RCE neural network is proved to be running fast and strong ability of anti-noise. Secondly, we use Freeman chain code to extract the distance from hand edge to the centre of the palm as feature vectors. Finally, we use those feature vectors as the input of RBF neural network and train the RBF neural network. Experiment results show this method is efficient and feasible. We develop a scissors-paper-stone game between human and humanoid robot using this method.4. In view of the massive data produced by vision system and large amount of processing, processing results are often delayed and the reaction of humanoid robot is slow. So a parallel computation model based on CUDA is introduced to improve performances of vision system.