A Study on Outdoor Guidance of Autonomous Land Vehicles by Binocular Computer Vision Based on Artificial Intelligent Policy

• Main Authors: yu-ching chang, 張煜青
• Other Authors: Rong-Chin Lo
• Format: Others
• Language: en_US
• Published: 2003
• Online Access: http://ndltd.ncl.edu.tw/handle/84844257100244703271

碩士 === 國立臺北科技大學 === 自動化科技研究所 === 91 === In this thesis, we have developed a new outdoor guidance system of autonomous land vehicle (ALV) equipped with binocular stereovision system based on artificial intelligence (AI) policy. The system uses two cameras to reconstruct the 3D structure of the scene. We can utilize the 3D information of the scene to recognize all obstacles. Hence, The ALV can perform navigation and obstacle avoidance in the outdoor environment using AI policy. The study topics in the thesis include camera calibration, stereo corresponding, improve information of the scene, obstacle avoidance and navigation based on AI policy. The correspondence problem is the important and most difficult problem of stereovision. The accuracy of 3D scene information will be affected greatly by the correspondence. In the study, first, we present a fast and simple approach to select reference points from the left image. The approach combines edge of the scene and several sensory vertical lines to select reference points. Then, we present a two-layer hierarchical stereo correspondence to the right image in binocular stereovision. A low-level processing is employed to obtain a set of points that are candidates for correspondence using Grey relation approach (GRA). The refinement of the set of low-level correspondences is performed by a high-level correspondence process. The high-level processing uses a permutation of genetic algorithm (GAP) approach for searching optimal solution from the solution space. Some constraints is embedded in the fitness function. The methods have been tested on a series of real images and perform very well. After the calibrated parameters and correspondence pairs are obtained, the 3D information of the scene can be computed by linear least-square method. But there are also several departures result from correspondence. We employ the nearest neighbor decision rules (NNDR) approach to solve this problem. Hence, we can obtain correct information of the scene via the approach on whole images. After we derive the 3D structure of the scene, we can know where the obstacles are. Hence, we propose navigation policy based on AI approach. The approach adopts a simple reactive navigation strategy by combining repulsion from obstacles with attraction to a goal. That is locally optimal in the sense of minimizing the hitting probability based on what is currently known about the world. Therefore, we can find an appropriate path for the ALV navigation. The ALV system has been performed in campus to demonstrate the effectiveness of the presented method.