【作者】 邵泽明；

【导师】 朱剑英；

【作者基本信息】 南京航空航天大学 ， 机械电子工程， 2009， 博士

【摘要】 随着移动机器人的研究和发展,其突出的工作性能在许多领域都被广泛应用,但同时也提出了更高的要求,越来越复杂的工作环境要求移动机器人智能化程度更高,自主能力更强,尤其是自主导航能力要更好。由于模拟人眼的视觉传感器技术具有较高的智能和优势,视觉移动机器人的发展越来越受到人们的重视,并表现出良好的发展前景,特别是在军事上的应用前景更加光明。本文针对视觉移动机器人在复杂、未知环境下的自主导航问题,重点研究了视觉定位、视觉障碍物检测以及视觉道路检测等关键技术,以提高移动机器人的智能水平,同时设计并实现了视觉移动机器人软、硬件系统。(1)视觉定位方面:研究了图像特征点的检测与匹配。针对图像特征点的检测,提出了分层快速SUSAN角点检测算法,利用角点周围像素的灰度特性,结合提升小波变换由粗到细的分层策略,快速找到角点特征,解决了SUSAN角点检测算法速度较慢的问题,为特征点的匹配打下基础。针对图像特征点的匹配,提出了一种新的RSTC不变矩特征点匹配方法,利用新构造的RSTC不变矩来度量角点的相似性,并用改进的RANSAC鲁棒估计以及外极线约束进行引导匹配,获得了较高精度的匹配结果,消除了野值匹配对所导致的长线条,较好地解决了图像特征点匹配不准确的问题。(2)视觉障碍物检测方面,研究了视差图计算方法,提出了自适应分层粒子群稠密视差图估计算法,首先利用SIFT特征检测与匹配算法准确确定视差范围;然后根据图像和视差范围的大小分层,建立由粗及细的自适应分层图像金字塔结构,加快搜索速度、减少错误匹配;最后在优化函数中引入根据匹配窗口大小自动变化的因子来调整灰度项和平滑项的权重,并用改进的带变异算子的整数形式的PSO进行优化,克服了遗传算法搜索的盲目性以及容易陷入局部最优的缺陷,能够更快、更好的找到最优解。(3)视觉道路检测方面,研究了图像处理中适合复杂图像处理的颜色模型,提出了SCT域归一化互信息道路检测算法,结合SCT颜色模型与人眼视觉系统的反应较接近、计算量小、抗噪能力强和归一化互信息测度鲁棒性较强、精确度较高的优点,充分利用像素点之间的关系进行加速处理,快速检测出移动机器人可行驶区域和可疑障碍物,为视觉系统障碍物的检测奠定基础。(4)在上述研究的基础上,本文设计并实现了视觉移动机器人的软、硬件系统。根据视觉移动机器人的功能需求,在“悍马”越野模型车的基础上,安装了左右伺服减速驱动电机、摄像机安装支架以及旋转机构,设计了与之相应的硬件控制系统,并搭配摄像机、无线影音传输模块以及图像采集卡等组成双目立体视觉系统的硬件部分,从而构成了视觉移动机器人平台;在硬件平台的基础上进行了视觉系统的开发,实现了视觉移动机器人自主导航控制功能。更多还原

【Abstract】 With the research and development of mobile robots, they are widely used, but more and more intelligence and autonomy are required to fit complex environment, so better auto navigation system is very important. As simulating human eyes, vision sensor technologies have more intelligence and advantages than other sensors, so more and more importance is added to vision mobile robots. Vision mobile robots have good development and future, especially in military. Auto navigation for vision mobile robots in complex and unknown environment has been focused on this dissertation. The key technologies of localization, obstacle detection and road detection based on vision technology for vision mobile robots have been researched to improve mobile robot’s intelligence, and a vision mobile robot system including hardware and software has also been established.In vision localization, image feature(corners) detection and matching method have been presented. As to image feature detection, a hierarchical fast corner detection algorithm by coarse-to-fine based on SUSAN algorithm has been proposed to improve runtime. According to the gray similarity around pixels and corner property in image, firstly the theory of lifting wavelet transform and coarse-to-fine hierarchical strategy have been used to find the coarse positions of corners, then SUSAN algorithm has been used to locate the corners accurately. As to feature matching, a corner matching method based on moment invariants of RSTC(Rotation, Scale, Translation and Contrast) has been proposed to improve precision and reliability. A new moment invariant of RSTC has been constructed to describe the corner features and to measure the similarity of corner matching, and a guided matching with improved RANSAC robust estimation and epipolar line constraint has been performed. The long lines caused by wrong feature matching have been eliminated, and better matching results has been got by this method than the gray matching method.In obstacle detection based on vision technology, a method of dense disparity map estimation using PSO algorithm with adaptive hierarchical images has been proposed to solve the stereo correspondence problem. In this method, firstly image features have been extracted and matched by SIFT algorithm, and the disparity range has been got easily and accurately. Then, according to restriction of the image size and the disparity range, the coarse to fine adaptive hierarchical image pyramids have been built to search fast and reduce wrong matching. With a regulation parameter varying with matching window used to give different power for grayness and smoothness data in optimization function while the matching window has been different in dissimilar supporting areas, an improved particle swarm optimization algorithm with variation operation for integer has been used to find the fittest solution from a set of potential disparity maps avoiding Genetic algorithm’s blind searching and easy getting in local best solutions. Experimental results on synthetic and real images have demonstrated that the proposed approach performs dense disparity estimation accurately and quickly.In road detection based on vision technology, road detecting using normalized mutual information of SCT color model has been presented to fit for complex images after comparing and analyzing different color models. The SCT color model is not only closer to HVS(Human Vision System), but also has the characters of strong noise-rejection and good real-time performance. Normalized mutual information has the advantage of good robustness and accuracy. They have been integrated in the proposed algorithm, and image processing time has been improved by using the relationship of pixels around. So safe areas and dubious obstacles information has been provided fast for vision system by the proposed algorithm.The hardware and software of vision mobile robot have been designed and implemented in this dissertation. Firstly, based on "Hummer" overland model car, two servo motors with planet type gearbox have been fixed on the left and right side for driving. Camera fixing bracket and rotary mechanism have been designed, and new hardware has also been designed for controlling. Moreover, with digital image acquisition card and the wireless video & audio transmitting system, the vision mobile robot platform has been built. Based on the platform, the software of vision system has been developed, and auto navigation by vision system has been implemented.更多还原