基于RGBD传感器的场景自适应性视觉里程计算法

摘要

For the problem of three-dimensional path estimation about mobile robots in unknown environment, a kind of calcula-tion method of scene adaptive visual odometry based on RGBD sensor was proposed. Firstly, the masking edge feature point, RGB edge feature points, feature point of ORB adaptively was selected by the texture information of the scene in the density. Sec-ondly, three equation systems were constructed by the KTL algorithm matching depth information of the corresponding feature points in the target frame and the reference frame. In order to improve the pose accuracy, the LM algorithm is used to minimize the projection error of the corresponding feature points. Finally the bundle adjustment is applied to optimize the robot’s overall trajectory. The experimental results show that, in the indoor scene with rich texture, the performance of the method is equivalent to RGBD SLAM, whereas in the scenes with sparse texture (FR2 slam, FR2 slam2) and no texture(FR3 str_notxt), the pro-posed method outperforms the RGBD SLAM algorithm. The offset A-RGBD SLAM algorithm of Transl. RMSE, Rot. RMSE, ATE RMSE were (54%, 66%, 54%;43%, 77%, 31%;60%, 43%, 81%) of RGBD SLAM algorithm.%针对移动机器人在未知环境下三维运动轨迹估计的问题,提出一种基于RGBD传感器的场景自适应性视觉里程计算法。该方法首先判断场景中纹理结构信息的疏密,自适应选择遮挡边缘特征点、RGB边缘特征点和ORB特征点,然后通过KTL算法匹配后,根据目标帧与参考帧对应特征点深度信息的有无,构建3组方程系统,再通过LM算法最小化对应特征点的投影误差来提高位姿精度,最后利用光束平差法来对机器人的运动轨迹进行整体优化。实验结果表明,在室内场景纹理丰富的条件下,该方法性能与RGBD SLAM相当,在纹理信息稀疏如FR2 slam, FR2 slam2和无纹理场景如FR3 str_notxt下该方法优于RGBD SLAM算法,其中A-RGBD SLAM算法的偏移均方根、旋转均方根、绝对轨迹误差分别只有RGBD SLAM算法的(54%,66%,54%；43%,77%,31%；60%,43%,81%)。