基于改进的3维ICP匹配的单目视觉里程计

针对目前流行的单目视觉里程计当移动机器人做“近似纯旋转运动”时鲁棒性不强的问题，从理论上分析了其定位鲁棒性不高的原因，提出了一种基于改进的3维迭代最近点（ICP）匹配的单目视觉里程计算法．该算法首先初始化图像的边特征点对应的深度值，之后利用改进的3维ICP算法迭代求解2帧图像之间对应的3维坐标点集的6维位姿，最后结合边特征的几何约束关系利用扩展卡尔曼深度滤波器更新深度值．改进的ICP算法利用反深度不确定度加权、边特征梯度搜索与匹配等方法，提高了传统ICP算法迭代求解的实时性和准确性．并且将轮子里程计数据作为迭代初始值，能够进一步提高定位算法的精度和针对“近似纯旋转运动”问题的鲁棒性．本文采用3个公开数据集进行算法验证，该算法在不损失定位精度的前提下，能够有效提高针对近似纯旋转运动、大场景下的鲁棒性．单目移动机器人利用本文算法可在一定程度上校正里程计漂移的问题．

Monocular Vision Odometry Based on the Improved 3D ICP Matching

The robustness of the current monocular vision odometries isn't strong enough when the mobile robot performs an approximately pure rotation. For this reason, a monocular vision odometry algorithm based on the improved 3D ICP (iterative closest point) matching is proposed by theoretically analyzing the reason of the low localization robustness. Firstly, the depth of the edge-feature point of the image is initialized. Then, an improved 3D ICP algorithm is used to iteratively solve the 6D poses of the 3D point sets between two frames. Finally, the depth value is updated by the extended Kalman depth filter based on the geometric constraint relation of edge-feature points. The improved ICP algorithm is superior to the traditional ones in the real-time performance and the accuracy, by using the inverse depth uncertainty weighting, the edge-feature gradient search, the matching and so on. What's more, it can further improve the localization accuracy and the robustness in case of the approximately pure rotation, by taking the wheel-mounted odometer data as the iterative initial values. The algorithm is verified on three public datasets, and it can effectively improve the robustness under the condition of the approximately pure rotation or in a large scale without decline of the localization accuracy. The proposed algorithm can correct the drift of odometer data for an actual monocular mobile robot to a certain extent.