| Autonomous Land Vehicle (ALV) is an intelligent mobile robot, which can run autonomously and continuously on road or cross-country. Its research, which involves theories and technologies of many disciplines, and embodies the latest achievements of information and artificial intelligence, is of great value in research and application, and is paid great attention to all over the world. Of all the ALV key technologies, vision-based navigation perceives and understands the surrounding quickly, and then determines a safe path for the ALV. The key to the vision-based navigation is the rapid detection and recognition of obstacles. Compared to other approaches, stereo vision is rapid, accurate and easy to conceal, as it is a passive measurement, which does not emit light and radiation. Therefore, stereovision navigation is a very important part of the ALV research. This thesis aims at the study and design of the stereovision system for the ALV.Perhaps the most important problem for stereovision is the accurate calibration of the cameras. Different from that in the laboratory environment, the calibration for the ALV must be done quickly and accurately as the cameras are mounted on the vehicle, and its procedure should be flexible as well. In the system that we have implemented, a flexible technique has been used to easily calibrate a camera. It is well suited for use without specialized knowledge of 3D geometry. The technique only requires the camera to observe a planar pattern shown at a few different orientations. Either the camera or the planar pattern can be freely moved. The motion need not be known. The test results prove that our approach is efficient and practical. In order to improve the computational efficiency of stereo matching by reducing the correspondence searching to ID (along the horizontal scan lines of image pairs), rectification is also a very important procedure. In this thesis, a simple and intuitional method, called two-step virtual rotation, is proposed. Furthermore, our new methods can be applied to a linear multi-baseline system.In chapter 3, we make a thorough study on algorithms of area-based stereo matching. Emphasis is put on both speed and robustness, as the final system is intended to work in a dynamic environment under real-time conditions. The topics of our research include choosing the matching criteria, window sizes, preprocessing and post-processing methods, and techniques of accelerating area-base stereo matching. The qualitative and quantitative analyses have been made based on the matching result of the synthesized and the real stereo pairs. In order to overcome the drawbacks of local algorithms, we propose a hierarchical stereo algorithm using dynamic programming. In our new algorithm, computation complexity is reduced by pyramid methods, and the ground control points are incorporated into the matching process to eliminate streak flaw of dynamic programming. In the end, the method of terrain reconstruction is also provided.It is well known that there are two difficult problems in the area of computer vision: (1) the algorithms of computer vision are far from mature; (2) real-time implementation of the algorithms is also very difficult. So the researches of computer vision fall into two categories: the algorithms and the implementations. Chapter 4, targeting at the intensive computational requirement and high transfer rate imposed by the stereo matching, probes into the architecture of real-time computervision. Its study covers the features of computer vision, methodologies, classification of hardware, architectural features for vision algorithms and computational resources. Finally, analyses and classifications of the procedures of the stereovision algorithms are made based on the research achievements, a three-layer hardware architecture for the stereo vision system consisting of FPGA-DSP-CPU is then proposed.Chapter 5 presents the implementation of the hardware paradigm. Detailed study of the hardware implementation of the procedures of algorithms is made. A...
|
PDF Full Text Request