【作者】 朱志刚；

【导师】 石纯一； 徐光祐；

【作者基本信息】 清华大学 ， 计算机应用， 1997， 博士

【摘要】 自然环境中视觉导航的研究涉及到计算机视觉中的各个主要方面，是一个有难度的综合性课题。视觉导航的基本任务是全局定位、道路跟踪和障碍物检测，其中每一部分都要以环境建模为基础。本文系统深入地研究了环境建模问题，包括： 1．提出了面向任务的多尺度和全覆盖视野的环境综合建模方法。将路标定位的全景图象建模，道路识别的全方位图象建模和目标（障碍物）检测的双目注视图象建模相结合，改进了依赖局部、单一信息的视觉导航方法。研究了面向任务的环境建模方法中的传感器设计、视觉算法和模型表示问题。 2．提出了由非精确摄象机运动下的图象序列建立3D环境全景模型的两步法，即基于运动滤波的图象稳定和基于时- 空－频域遮挡模型的全景外极面图象分析。本文推广了全景图象方法和外极面图象方法，从而使之能适用于具有抖动的图象序列分析，并避免了一般运动视觉方法的不适定问题和特征对应问题、基于空域约束的迭代方法的局部最小化等问题。 3．提出了全方位图象特征和神经网络相结合的道路建模方法，较好地解决了机器人依赖局部视野信息迷路的问题、视觉算法依赖特定环境特征推广性差等问题。本文提出了主分量分析和 Fourier变换相结合的全方位图象数据压缩和旋转不变特征提取方法，设计了在识别道路类型基础上进行道路方向估计的组合神经元网络，从而提供了解决机器人在不同类型道路上自动切换和自适应道路跟踪的可行方法。 4．提出了基于重投影变换的障碍物检测方法，跳出了传统立体视觉特征抽取、匹配和三维恢复的模式。设计了无特征提取和对应的立体视觉新算法，给出了克服摄象机俯仰影响的动态重投影变换算法，增强了系统在颠簸的道路上运行的适应性。利用重投影变换后双目图象对上路面特征零视差的特性，通过障碍物有无判断和障碍物三维测量的“分步渐进”过程，可高效、可靠地对路面障碍物实时检测。 5．系统实现：门）实现了全景建模中的图象稳定、遮挡恢复和深度分层，从而为全局定位的自然路标提取和真实环境再现的图象合成打下了基础；C）设计和实现了适合于室外环境的全方位成象系统和单摄象机双目立体成象系统：m实验验证了全方位道路图象的神经网络学习方法的有效性；w实现的实时障碍物检测系统己经过大量的室外道路环境下障碍物检测的试验，结果说明系统具有很强的实用性。更多还原

【Abstract】 Visual navigation of a mobile robot in the natural environment is a difficult and comprehensive subject which is related to almost every aspects of computer vision researches. The fundamental tasks of visual navigation are composed of global localization, road following and obstacle detection. Environment modeling is the foundation of visual navigation. This dissertation is devoted to the deep and systematic study on the visual environment modeling. The main contributions include: 1. A task-oriented, multi-scale and full-view visual modeling strategy is proposed for the natural environment which combines the panoramic vision for scene modeling, omni-directional vision for road understanding and binocular vision for obstacle detection together. This approach overcomes the drawbacks of traditional visual navigation methods that mainly depended on local and/or single view visual information. In this direction sensor design, data processing and model representation are closely explored. 2. A two stage method is presented for the 3D panoramic scene modeling from vibrated image sequences which consists of (I) image stabilization by motion filtering and (2) depth estimation and depth boundary localization. The two stage method not only combines Zheng and Tsuji抯 panoramic image method with Bakers epipolar plane image analysis, resulting the so called panoramic epipolar plane image method, but also generalizes them to handle image sequence vibrations due to the un- controllable fluctuation of the camera. The two stage method by-passes the correspondence problem and ill-posed problem encountered in the general motion analysis, and avoids the local minimum problem of the spatial-constrain-based iteration method. 3. .A new road following approach . the Road Omni-View Image Neural Networks (ROVINN). is proposed which combines the omni-directional image sensing technique with neural networks. The ROVINN makes the robot never get lost and enables it to learn from the road images. The ROVINN approach brings Yagi抯 COPIS to the outdoor road scene and provides a new solution from the CMU抯 ALVINN. Compact and rotation-invariant image features are extracted by integrating the principle component analysis (PCA) and the Fourier transform (DFT). The modular neural networks can estimate road orientations more efficiently by first classif~抜ng the roads, and thus enable the robot to adapt to various road types automatically. 4. A novel method called image reprojection transformation is presented for road obstacle detection based on binocular vision. Dynamic reprojection transformation algorithms are developed so as to work in un-even road surface. The novelty of the (dynamic) reprojection transformation method, which ensembles the gaze control of the human vision, lies in the fact that it brings the road surface to zero disparity so that the feature extraction and matching procedures of the traditional stereo vision are avoided in the obstacle detection task. The progressive processing strategy of reproj ection transformation, yes/no verification and obstacle measurement make the obstacle detection efficient, fast and robust. 5. System implementations.(l) In the 3D panoramic scene modeling , the algorithms of motion filtering and image stabilization, kinetic occlusion detection and depth layering, have been developed so as to found a ground base for landmark selection of global localization and image synthesis of virtualized rea更多还原