基于直接法的机器人半稠密地图构建研究

发布时间：2018-07-15 11:51

【摘要】：机器人同时定位与地图构建技术是全自主机器人的关键技术。虽然,基于单目视觉的同时定位与地图构建方法相较于基于激光的有价格低、能够构建三维地图等优点,但其自身存在计算量大、易产生尺度漂移、建图精度不高等问题。因此,如何降低基于单目视觉的同时定位与地图构建算法计算量,提高地图构建精度已成为当今机器人地图构建领域热点。针对基于单目视觉同时定位与地图构建方法所存在的计算量大、易产生尺度漂移、建图精度不高的问题,通过采用基于相似变换的闭环检测与改进直接法,结合优化方法,重点开展了半稠密地图构建算法研究,并通过实验验证了本文所建立算法的正确性与有效性,提高了单目视觉同时定位与地图构建精度,降低了计算量。具体进行了如下研究:首先,针对如何提高基于直接法地图构建的快速性问题,开展了基于Lucas-Kanade方法的相机位姿跟踪改进方法研究。结合机器人行走速度指标,基于Lucas-Kanade方法,利用雅克比矩阵对该方法的迭代过程进行了分析,获得该方法计算量大的原因,在此基础上,通过位姿变换增量矩阵的逆实现了对Lucas-Kanade方法的改进,利用TUM数据集实验验证了改进LucasKanade的有效性;通过实验分析了地图稠密度对相机跟踪算法快速性的影响,实验中获得了算法的时间代价随像素梯度阈值变化曲线,得到了兼顾地图稠密程度和相机跟踪算法快速性的像素梯度阈值。然后,针对基于单目的地图构建算法尺度漂移和位姿估计不准确问题,开展了基于直接法的闭环检测和地图优化研究。借鉴了基于特征点法解决单目SLAM尺度漂移问题的方法,将相机的位姿变换和场景尺度同作为优化变量求解关键帧之间的相似变换,通过最小化像素灰度和深度残差平方和求得相似变换的最优估计。在此基础上,将求得的相似变换作为优化变量,通过求解最小化代价方程实现地图的全局优化。利用TUM数据集实验验证了基于相似变换的闭环检测和地图优化的有效性最后,为验证算法的快速性和准确性,在室内实验中以激光估计的轨迹为参照轨迹计算算法准确性,实验中对比各算法时间代价以验证算法快速性;为验证算法在解决单目SLAM尺度漂移问题的有效性,室外实验中手持相机进行长距离的闭环行走,对比优化前后的地图,验证了基于相似变换的地图优化在解决单目SLAM尺度漂移问题的有效性。
[Abstract]:The technology of simultaneous localization and map building is the key technology of autonomous robot. Although the method of simultaneous location and map construction based on monocular vision has the advantages of low price and ability to construct 3D map compared with that based on laser, it has many problems, such as large amount of calculation, easy to produce scale drift and low precision of mapping. Therefore, how to reduce the amount of computation of simultaneous localization and map construction based on monocular vision and improve the accuracy of map construction has become a hot spot in the field of robot map construction. Aiming at the problems of large computation, easy to produce scale drift and low precision of map building based on monocular vision simultaneous localization and map construction, this paper adopts close-loop detection and improved direct method based on similarity transformation, and combines optimization method. The research of semi-dense map construction algorithm is emphasized, and the correctness and validity of the proposed algorithm are verified by experiments. The accuracy of monocular vision simultaneous location and map construction is improved, and the computational complexity is reduced. The specific research is as follows: firstly, aiming at how to improve the rapidity of map construction based on direct method, the improved method of camera pose tracking based on Lucas-Kanade method is studied. Based on Lucas-Kanade method, the iterative process of this method is analyzed by using Jacobian matrix, and the reasons for the large amount of calculation are obtained, based on Lucas-Kanade method. The Lucas-Kanade method is improved by the inverse of the incremental matrix of pose transformation, the effectiveness of the improved LucasKanade is verified by the TUM dataset experiment, and the influence of the density of the map on the fast tracking algorithm is analyzed. In the experiment, the time cost of the algorithm varies with the pixel gradient threshold, and the pixel gradient threshold, which takes into account the density of the map and the rapidity of camera tracking algorithm, is obtained. Then, aiming at the inaccuracy of scale drift and pose estimation based on single-destination graph construction algorithm, the closed-loop detection and map optimization research based on direct method are carried out. Based on the feature point method, the camera pose transformation and scene scale are used as optimization variables to solve the similarity transformation between the key frames. The optimal estimation of similar transformation is obtained by minimizing the sum of gray and depth residuals of pixels. On this basis, the obtained similarity transformation is regarded as the optimization variable, and the global optimization of the map is realized by solving the minimization cost equation. The validity of close-loop detection and map optimization based on similarity transformation is verified by TUM dataset experiment. Finally, in order to verify the speed and accuracy of the algorithm, the laser estimated trajectory is used as the reference trajectory calculation accuracy in indoor experiments. In order to verify the effectiveness of the algorithm in solving the problem of single-eye slam scale drift, the hand-held camera carries out long distance closed-loop walking in outdoor experiments, and compares the map before and after optimization. The validity of map optimization based on similarity transformation in solving the problem of single scale drift of slam is verified.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242

【参考文献】