基于双目视觉的六足机器人环境地图构建及运动规划研究

发布时间：2018-05-29 02:49

本文选题：六足机器人 + 双目视觉　；参考：《哈尔滨工业大学》2016年博士论文

【摘要】：六足机器人与其他轮式移动机器人相比,具有较强的地形适应能力和良好的运动稳定性,在军事侦察、抢险救灾、星际探索、反恐爆破、考古探测等多个领域都开始展现其作用。六足机器人在适应复杂环境、独立完成作业方面的能力是亟待解决的关键问题。课题研究将以解决六足机器人复杂环境自主运动问题为出发点,重点攻克实时性环境建模、自主定位及运动规划三个基础理论难题,最终以典型复杂作业环境应用为示范,验证所取得的提高机器人复杂环境下自主运动能力的理论与技术创新成果。为解决立体匹配中匹配速度与匹配精度相互制约的问题,提出一种基于支撑点的立体匹配算法。首先,对摄像机采集到的原始图像进行极线校正,确保待匹配图像对的对应特征点在同一极线上。其次使用Canny边缘检测算子求取摄像机左图像的边缘点作为稳定支撑点,应用分治法对支撑点集进行2D Delaunay三角网格划分,以剖分后的三角形面片作为视差匹配基元,建立视差模型和能量最小化函数进行初始视差估计。最后根据三角形网格共用顶点的特性,对初始视差进行修正,获得完整的视差图,经重投影矩阵恢复出场景的三维信息。提出基于离散点云三角剖分的地形三维重建方法,为六足机器人运动规划提供前方环境信息。此方法将整个重建过程分为三个步骤:采用空间包围盒算法对曲面离散点集进行划分,加速离散点k邻近点的查找,参数化点云数据后构造其切平面,采用基于曲率的方法对初始点云进行数据精简,有效降低了点云拓扑结构的复杂性;点集划分完成后,区域内根据异侧不交约束、法向量夹角最大约束、半圆距离约束和最小内角最大约束进行Delaunay三角剖分;根据三角形的重数进行各剖分区域的融合与拼接,实现最终三角剖分。该算法适用于闭合曲面和非闭合曲面离散点云的三角剖分,剖分后的三角网格均匀、平滑,能较好的体现曲面的细节特征。算法执行速度快,近万个离散点云的剖分时间仅为5.8s。针对移动机器人仅通过双目视觉来进行定位构图存在精度低和鲁棒性差等问题,提出一种基于GPU加速的SIFT特征匹配算法,实现了六足机器人的同步定位与地图构建。应用基于GPU加速的SIFT算法检测机器人运动前后帧图像匹配的特征点对,根据迭代邻近点算法计算其旋转与平移,实现机器人的位姿估计。为了防止因场景照明不良或者机器人运动不稳造成图像模糊,导致视觉信息丢失造成视觉里程计失灵的问题,结合AHRS传感器实现六足机器人的SLAM。该方法既可以解决单目视觉利用特殊初始化方法获取环境特征点信息不准确的问题,也可以避免单一使用双目视觉SLAM恢复运动带来的计算量大的缺点。实验结果表明,在室内环境下,算法运行稳定,定位精度高。在前述基于双目视觉系统采集地形信息并进行地形构建的基础上,依据地形的几何信息和六足机器人自身运动能力进行落足点的初步选取。地形的几何信息主要考虑到其崎岖度、高度及可落足区域的面积。机器人的运动能力主要依据单腿的可达能力和机器人位姿的稳定性。在可落足点较多的情况下,依据机器人的稳定裕度及灵活性进行落足点的筛选与确定。提出基于障碍物最小凸包曲线的六足机器人摆动相足端轨迹规划算法,在降低机器人的能耗的同时提高了运动效率。提出机器人的位姿调整算法,通过调整机器人的重心位置、ZMP点至躯干支撑多边形最大内切圆的圆心处,实现机器人在平坦地形及崎岖地形的位姿调整,保证了六足机器人下一步运动的稳定性与灵活性利用六足机器人平台HIT-II开展综合实验研究,设计并实现了六足机器人系统的控制程序。通过三组实验验证了针对六足机器人未知环境下自主行走所提方法和理论的有效性。第一类实验场景为平坦地形,包含凸起和凹陷障碍。通过分析环境感知时间、感知精度及机器人的位姿变化,验证了机器人在该类地形的通过性能。第二类实验为柱状地形,重点验证了落足点选取算法和位姿调整算法的有效性。第三类实验场景为仿丘陵地形,通过立体匹配算法恢复出地形的点云信息,依据三角剖分算法重建出机器人前方环境地图。通过分析机器人行进过程中的位姿变化,验证立体匹配算法、落足点选取算法及位姿调整算法的可靠性和有效性。过实验对比发现,地形的复杂程度影响机器人的落足点选取及足端轨迹规划的执行效率,进而影响机器人的行进速度。六足机器人系统的综合实验充分验证了所提立体匹配算法、地形构建算法、落足点选择算法、足端轨迹规划算法及位姿调整算法的有效性及可靠性。
[Abstract]:Compared with other wheeled mobile robots, six legged robots have strong terrain adaptation ability and good motion stability. They are playing an important role in many fields, such as military reconnaissance, emergency rescue and disaster relief, interstellar exploration, anti-terrorism blasting, archaeological detection and so on. The ability of six foot robots to cope with the complex environment and to complete the work independently is urgent. The key problem to be solved is to solve the problem of autonomous motion of complex environment of six legged robots as the starting point, focusing on three basic theoretical problems of real-time environment modeling, autonomous positioning and motion planning, and finally taking the typical complex operation environment as demonstration to verify the improvement of autonomous motion under the complex environment of the robot. In order to solve the problem of mutual restriction between matching speed and matching precision in stereo matching, a stereo matching algorithm based on support points is proposed. First, the polar line correction of the original image collected by the camera ensures the corresponding feature points in the matching image on the same pole line. Secondly, the Canny is used. The edge detection operator takes the edge point of the left image of the camera as the stable support point, and uses the divide and conquer method to divide the 2D Delaunay triangular mesh of the support point set. The triangle face after subdivision is used as the parallax matching base, and the parallax model and the energy minimization function are established for the initial parallax estimation. Finally, the triangle mesh is shared by the triangular mesh. The characteristics of the vertex, the correction of the initial parallax, the complete parallax graph, the restoration of the three-dimensional information of the scene through the reprojection matrix. A three-dimensional reconstruction method based on the discrete point cloud triangulation is proposed, which provides the front environment information for the motion planning of the six legged robot. This method divides the whole reconstruction process into three steps: the use of space The bounding box algorithm divides the discrete point set of the surface, accelerates the search of the adjacent points of the discrete point K, constructs its tangent plane after the parameterized point cloud data, and uses the curvature based method to simplify the data of the initial point cloud, effectively reducing the complexity of the topological structure of the point cloud. After the point set is completed, the normal vector is based on the non side constraint and the normal vector. The maximum constraint of the angle, the semicircle distance constraint and the minimum inner corner maximum constraint are Delaunay triangulation, and the final triangulation is realized by the fusion and splicing of the subdivision regions according to the weight of the triangle. The algorithm is suitable for the triangulation of the closed surface and the discrete point cloud of the closed surface. The triangular mesh after the split is uniform, smooth and capable. The algorithm executes the details of the surface better. The algorithm performs fast, and nearly 10000 discrete point clouds are divided by 5.8s. only for the problems of low precision and poor robustness for mobile robots only through binocular vision. A SIFT feature matching algorithm based on GPU acceleration is proposed to realize the synchronization of six legged robots. Using the SIFT algorithm based on GPU acceleration to detect the feature points of the frame image matching before and after the motion of the robot, and calculate its rotation and translation according to the iterative neighborhood point algorithm, and realize the position and pose estimation of the robot. In order to prevent the image blurred caused by the scene illumination or the robot movement instability, the visual information is lost. As a result of the failure of the visual odometer, the implementation of the SLAM. method by combining the AHRS sensor with the AHRS sensor can not only solve the problem of obtaining the inaccurate information of the environmental feature points using the special initialization method in monocular vision, but also avoid the shortcoming of the large amount of computation brought by the single visual SLAM recovery motion. The experimental results show that In the indoor environment, the algorithm runs steadily and the positioning accuracy is high. Based on the previous biocular vision system collecting terrain information and constructing the terrain, the initial selection of the landing point is carried out according to the geometric information of the terrain and the motion ability of the six foot robot. The geometric information of the terrain mainly takes into account its rugged, height and fall. The area of the foot area. The motion ability of the robot is mainly based on the ability of the single leg and the stability of the robot position. In the case of more landing points, the landing point is selected and determined according to the stability margin and flexibility of the robot. The foot track gauge of the wobble phase of the six foot robot based on the minimum convex hull line of the obstacle is put forward. The calculation method improves the motion efficiency while reducing the energy consumption of the robot. The position and posture adjustment algorithm of the robot is proposed. By adjusting the center of gravity of the robot and the ZMP point to the center of the maximum inner tangent circle of the trunk support polygon, the position and posture adjustment of the robot in the flat terrain and rough terrain is realized, and the next step of the six foot robots is ensured. The dynamic stability and flexibility of the six legged robot platform HIT-II are used to carry out comprehensive experimental research. The control program of the six legged robot system is designed and realized. The effectiveness of the method and theory of the autonomous walking for the six foot robot is verified by three groups of experiments. The first class experiment scene is a flat terrain with a convex hull. By analyzing the time of environmental perception, the accuracy of perception and the change of the position and posture of the robot, the performance of the robot in this kind of terrain is verified. The second class experiment is a columnar terrain, and the validity of the algorithm for selecting the landing point and the algorithm of position and posture adjustment is verified. The third kind of actual scene is a hilly terrain, and the stereo matching is calculated by the stereo matching. This method restores the point cloud information of the terrain, and reconstructs the environment map in front of the robot according to the triangulation algorithm. By analyzing the change of the position and posture in the process of robot moving, the stereo matching algorithm, the landing point selection algorithm and the position and attitude adjustment algorithm are reliable and effective. The complexity of the terrain is found to affect the robot. The selection of the landing point and the execution efficiency of the trajectory planning of the foot end affect the moving speed of the robot. The comprehensive experiments of the six foot robot system fully verify the validity and reliability of the proposed stereo matching algorithm, the terrain construction algorithm, the landing point selection algorithm, the foot trajectory planning algorithm and the position and attitude tuning algorithm.
【学位授予单位】：哈尔滨工业大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41;TP242

【相似文献】