A Study On Obstacle Avoidance Strategy For Humanoid Robot Based On Monocular Vision

With rapid development of related disciplines of artificial intelligence, humanoid robots become one of the most popular robotics research. The ability to walk in human habitat is needed for humanoid robots to be widely applied in human daily life, and the most important aspect of walking is avoiding obstacles during walking. This paper mainly studies autonomous obstacle avoiding of humanoid robot, which relates to robot vision and path planning. Monocular vision is a single camera-based vision system and is widely used as a solution for environment perception of robots. Most of path planning methods for humanoid robots are based on robot vision system, which provides envi-ronment information, mainly relative locations of destination and obstacles to robots, with which a collision-free path to destination can be designed.The main work of this paper:1. An obstacle recognition method for the single-camera vision system is pro-posed. At first, some reasonable assumptions are made about the environment the robot will walk on. Then, obstacle regions are obtained based on color features and image segmentation. According to color features of floor samples in HIS color space, pixels in real-time images are classified into belonging to obstacles and non-obstacles. Binary image is obtained by image binaryzation and morphology processing and locations of all obstacles are estimated in it with image segmentation. At last, some rectangles sur-rounding the obstacles are calculated and used to represent the obstacle regior.s. Based on region growing, image segmentation is implemented effectively with a data structure queue and label matrix. The parameters in all digital image processing methods used in this paper are obtained by some experiments and they are keys for achieving a good recognition result.2. With the obstacles being recognized, an obstacle localization method is pre-sented for single-camera vision system. At first, a pin-hole imaging model of single-camera is created and the coupling of the model is analyzed. Then, the camera is cal-ibrated to obtain the mapping between pixel coordinate and real-world coordinate of the feature points of the obstacles. On the basis of SVM and mean field theory, a new camera calibration method with SVM learning accelerated by mean filed theory is pro-posed. At last, a training-testing scheme based on corner detection is constructed to calibrate the camera and camera calibration toolbox in MATLAB is used to verify the accuracy of the proposed method. In order to enhance the ability and speed of corner detecting, traditional Harris corner detection algorithm is improved with pretreatment method based on similarities of8-neighbourhood. 3. A real-time online tree-like path planning algorithm is proposed to make hu-manoid robot walk free on ground with obstacles. At first, some exploratory points located in a semicircle around the robot are generated and one of them is chosen as the temporary destination according to the relative orientation and distances of destination and obstacles. Then, a fuzzy controller is designed to determine the proper walking mode to reach the temporary destination. The procedures below are repeated to make the robot reach the final goal in a collision-free path. At last, the proposed method is implemented and verified on a physical platform.