【作者】 周自维；

【导师】 赵杰；

【作者基本信息】 哈尔滨工业大学 ， 机械电子工程， 2014， 博士

【摘要】 移动机器人对外部环境的有效感知以及基于该感知信息的路径规划是实现其自主运动的前提和关键问题。双目立体视觉不但能够感知环境的影像信息，而且能够通过两幅视图的匹配关系构建场景三维数据，是获取场景三维空间信息的理想方式。六足机器人与其他类型的移动机器人相比，具有较强的地面适应能力和良好的运动稳定性。本文在六足机器人基础上研制双目立体视系统，使之具备三维信息感知能力，并且对六足机器人系统在复杂环境下的路径规划方法进行研究，具有重要的理论和实际应用价值。本文在对双目立体视觉构建原理进行深入分析的基础上，以六足机器人负载能力和应用特点为指标，设计小巧紧凑的双目立体视觉机构。完成以DSP为核心、以CAN总线为框架的分布式控制单元，实现了立体视觉与六足机器人系统的一体化设计，构成完善的具有三维信息感知能力的六足机器人系统。在此基础上对双目立体视觉系统使用的摄像机进行单目和双目的标定与校正，为后续的立体匹配研究提供理想的视图对和摄像机参数信息。立体视觉匹配研究中，采用全局最优匹配方法能够获取高精度视差图，本文针对全局最优立体匹配算法求解时间过长的问题进行深入研究，设计了两类立体匹配算法提高匹配速度。第一类匹配算法建立基于马尔科夫随机场最大后验概率模型的全局最优目标函数，采用最信度传播算法求取最优解。为进一步提高全局最优算法的求解速度，设计基于最优分割的信度传播改进算法。采用最优分割策略将视图分割为若干小的区域，将全局问题局部化，运用OpenMP多核加速技术对多个分割区域进行并行匹配，提高整体视图的匹配速度。通过最优分割与并行处理，算法的匹配时间减少到3秒以内，综合匹配精度达到91%；在第二类匹配算法中设计支撑点扩展快速匹配算法，采用形式更加简化的立体匹配目标函数进行匹配。算法首先搜索出具有明确匹配关系的点集并命名为支撑点，在整个视图空间的匹配中，以支撑点为核心向邻域扩展，在扩展区域的匹配过程中采用全局次优快速匹配算法。由于扩展空间以支撑点为邻域，因而利用支撑点的匹配结果为已知条件，有效缩减待待匹配区域的视差范围，进而提高匹配速度。算法匹配时间达到0.6秒以内，综合匹配精度达到78%，匹配速度有明显改善，能够达到实际应用的需求。以视差图为基础经过三维重投影恢复场景三维信息，采用基于隐式马尔科夫随机场模型识别方法对场景三维区域进行划分，将场景划分为通行区域与障碍区域，为机器人运动规划提供环境基础。针对局部信息路径规划问题，提出双层路径规划算法。双层路径规划算法由两个层次组成，分别是路径规划层和运动控制层。路径规划层中的快速搜索多边形构造方法能够在复杂环境下搜索出逃离障碍物的包围多边形，构建可视图，通过可视图优化获取最短路径；运动控制层采用“基于势场函数的机器人运动控制算法”，控制机器人实际运动，使机器人能够有效躲避障碍物并逐步趋向目标点。每个控制周期内，规划算法的两个步骤循环运行，控制机器人完成规划运动。为体现本文两类立体匹配算法的作用，设计了结合两类匹配算法优点的综合感知实验方法。综合感知方法对场景信息的获取通过两个步骤实现，首先采用快速支撑点扩展算法进行立体匹配，恢复场景三维信息，将场景分为通行区域和障碍区域。为保证通行区域内机器人运动的安全性，利用视差图与场景三维点的对应关系，采用高精度匹配算法对通行区域重新匹配以获取更精确的场景信息。在六足机器人路径规划实验研究中，设计六足机器人步行方法，该六足机器人步行方式与路径规划算法输出的结果相对应，进而验证路径规划算法的实际价值。实验中设计了平坦地面与粗糙地面两类实验环境，在两类环境中采用综合感知法、高精度感知法和快速感知法分别完成实验。实验结果表明，综合感知法能够获取更加理想的实验效果。六足机器人系统整体实验也充分验证了本文提出的三维构建算法、地面识别算法的实际效果和应用价值。更多还原

【Abstract】 During the research of robot autonomy movement, the core subject is how to perceive the environment and decide their own motion planning. We design and realiz a binocular stereo vision mechanism system firstly. The mechanism of binocular stereo vision system has compact design, its electrical control system is designed based on distributed framework with CAN bus, in which DSP processor acts as the core. The motion controler uses servomotor as driver, which has high precision and flexibility. The binocular stereo vision mechanism system and the hexapod robot make up a perfect hexapod robot stereo vision platform. To acquire the intrinsic and external parameters of the cameras used in binocular system, we achieved monocular and binocular camera calibration respectively and provide ideal stereo image pairs and camera parameters for stereo matching study, the matching results lay the foundation for three-dimension scene construction.The disparity map is the basic for building three-dimensional information, and the matching algorithm is the hot point in researches on stereoscopic vision. Although local optimization stereo matching algorithm is easy to implement, it is usually has low quality which difficult to apply in real application. The global optimization algorithm, which building a high quality disparity map, is an emphases in the research of stereo matching. Currently available globally optimal algorithms can get high precision disparity map, but the matching time is relatively long and need several minutes usually.Two types of stereovision matching algorithm with high precision are designed in this paper. The first algorithm bases on MRF/MAP and use maximum accumulation rule of belief propagation algorithm to obtain disparity. In order to improve the matching speed, an global optimal division algorithm is designed. The algorithm uses optimal segmentation strategy to divide the whole image into a number of small areas, matching each independent split region using OpenMP acceleration tools to reduce overall disparity map matching time. Experimental results show that using optimal segmentation and parallel processing, the matching time of the improved belief propagation algorithm reduce to less than3seconds with overall accuracy of91%. Although global segmentation algorithm accelerate the matching time, it is hard to use in practical application still. To achieved real-time disparity matching, the support points expansion algorithm is designed to speed up the matching time. The algorithm redesigns simple objective function firstly. During the searching of optimal solution, design the support points expansion method. The matching points with a clear matching relationships, which named as support points are searched out. In subsequent step, expansion the neighborhood of the support points and matching the extended space. For the support point give the prior knowledge to the neighbor points, it reducing global solution space, thereby improving the matching speed. According to the results of the experiment, the matching time less than0.5seconds, the whole match accuracy up to78%, although its match quality is relatively lower than the global segmentation algorithms, the matching speed has improved greatly, which satisfied application requirements.After acquire disparity map, the three dimensional scene is recovered. The obstacle recognition method based on Hidden Markov Model is used to classify the scene, which supports environmental information for robot autonomy movement. In the research of robot autonomous motion, design a apposite way polygon construction algorithm and robot motion controller based on potential field function. Apposite way polygon construction algorithm can construct a surrounds polygons of the barrier in a complex environment, generate optimal paths pass through obstacles. Robot motion controller set attracting potential field function and repulsive potential field function as input vector, calculate speed and angle of the robot as output vector. The controller controls the robot movement, so that robots can effectively avoid obstacles and move towards target point; controller set the basic speed of the robot to solve local minimization problem caused by zero potential field. Experiments based on artificial map showed that compared with the artificial potential field-APF and wall-following–Bug, our method acquire better optimization path and real-time performance, with a wide range of practical applications.To reflect the value of the two matching algorithms in hexapod robot experiment, designed a comprehensive perception method. Comprehensive perception method composes with two steps. Using rapid stereo matching method to restore the entire scene3D points, with hidden Markov models to identify passable area. In order to ensure the safety of the robot in passable area, according the corresponding relationship between disparity map and3D points, using the high quality matching method to identify the passable area again to obtain more precise scene information. Experimental results show that compare with fast matching method or high quality matching method alone, comprehensive perception method is more effective in the total path length and total exercise time.To verify the value of our autonomous algorithm, rough terrain recognition algorithm and stereo matching algorithm, the final hexapod robot experiments are designed. Two types of ground is designed namely flat ground and rough ground. The experiments result show that the proposed algorithms have research value on the robot practical application.更多还原