基于改进视觉里程计算法的移动机器人定位

Localization of Mobile Robot Based on Improved Visual Odometry

针对自主移动机器人定位的问题,该文提出一种改进视觉里程计算法。存在动态物体的环境中,传统的视觉里程计定位误差较大。为有效解决此问题,该文提出一种面向动态环境的基于点面特征的视觉里程计算法。在实时性得到保证的前提下,该算法可以有效降低场景中动态物体对位姿估计的影响,提高机器人定位精度。首先,基于相对距离不变原则,算法将运动状态相同的ORB特征点聚类;然后,针对环境中平面较多的特点,利用DBSCAN+KDTREE算法对点云局部法向量聚类,得到场景平面参数;最后,根据得到的相互垂直的平面数量的不同分类计算相机位姿变换矩阵。TUM的RGBD数据集与实际场景测试评估表明,所提算法取得了良好效果。

Aiming at the problem of autonomous mobile robot localization,an improved visual visual odometry is proposed.In scenes with dynamic objects,the traditional visual odometer has a large positioning error.In order to solve this problem effectively,a visual odometry based on the features of points and planes in a dynamic environment is proposed.Under the premise of ensuring real-time performance,the algorithm can effectively reduce the impact of dynamic objects in the scene on pose estimation and improve the localization accuracy of the robot.First,Based on the principle of relative distance invariance,ORB feature points with the same motion state are clustered by the algorithm.Then,in view of the environment with more planes,the DBSCAN+KDTREE algorithm is used to cluster the local normal vectors of the point cloud to obtain the scene plane parameters.Finally,The algorithm calculates the camera pose transformation matrix according to the different classifications of the number of mutually perpendicular planes.TUM’s RGBD data set and actual scene test evaluation show that the proposed algorithm has achieved good results.