平面运动约束下的相对位姿估计最小配置解

Minimal solver for relative pose estimation under planar motion

基于视觉的相对位姿估计是自动驾驶车辆等移动机器人进行自主定位与环境感知的核心关键技术.为提高相对位姿估计算法的精度、效率和鲁棒性,寻求相对位姿估计的最小配置解一直是重点关注的研究热点.传统的相对位姿估计算法通常仅使用图像匹配点对中的图像坐标信息,忽略特征描述子提供的特征旋转角和尺度等额外信息.针对移动机器人应用中单目相机平面运动这一常见情景,本文利用图像特征描述子提供的信息中的约束方程,提出了一种新的平面运动估计最小配置解.根据地面场景上图像特征满足的单应约束,利用单个旋转不变特征中的图像坐标和旋转角信息,闭式求解单目相机平面运动的相对位姿.由于该方法所需要的特征数量最少,因此可以高效地与RANSAC方法或直方图投票方法相结合,用于两视图之间的初始运动估计和误匹配点对剔除.仿真和公开数据集实验表明,本文能够明显提高单目相机平面运动估计的精度和鲁棒性,可运用于自动驾驶车辆等移动机器人在地面、道路等典型场景上的自主定位与视觉感知.

Visual-based relative pose estimation is a core technology for autonomous localization and environment perception of mobile robotssuch as autonomous driving vehicles. In order to improve the accuracy, efficiency, and robustness of the algorithm, the minimalsolvers for relative pose estimation are an important research topic. Traditional relative pose estimation algorithms typically utilizeonly the image coordinate information of the matched feature points, which ignores the additional information provided by featuredescriptors, such as feature rotation angles and scales. In this paper, focusing on the common scenario in mobile robot applications,we propose a new minimal solver for planar motion estimation using the constraints provided by the feature descriptors. By exploitingthe homography constraints, we obtain a closed-form solution for the relative pose of the monocular camera utilizing a single rotationinvariantfeature. Due to the minimal number of features required by the proposed solver, it can be efficiently combined withRANSAC or histogram voting methods for initial motion estimation and removal of outlier matches. Experiments on synthetic dataand public datasets demonstrate that our method significantly improves the accuracy and robustness of monocular camera planarmotion estimation, which makes it applicable to autonomous localization and visual perception of mobile robots.