Em-Based Point to Plane ICP for 3D Simultaneous Localization and Mapping
Em-Based Point to Plane ICP for 3D Simultaneous Localization and Mapping
复制标题
DOI：
10.2316/journal.206.2013.3.206-3806
复制
发表时间：
2013
期刊：
影响因子：
--
通讯作者：
Yue Wang;R. Xiong;Qianshan Li
中科院分区：
文献类型：
--
作者：
Yue Wang;R. Xiong;Qianshan Li
三维同时定位与地图构建（SLAM）是自主机器人研究的一个重要课题.作为SLAM前端应用的流行算法之一是迭代最近点（ICP）。在本文中，ICP建模成一个概率框架，包括姿态估计和数据关联步骤，使用期望最大化（EM）。导出的结果是，如果姿态估计和数据关联步骤都采用相同的度量，则解收敛到局部最小值。因此，决定度量形式的度量模型应该是算法的关键因素。然后，以测量模型的形式分析了点到点、点对面和面对面的测量模型，揭示了它们对两次扫描之间连接的描述。在分析的基础上，提出了一种改进的点-面测量模型，利用特征值分解估计各平面的协方差，放松ICP的模型假设，从而获得更好的解。实验结果表明，该算法具有较好的性能，与理论结果相一致。
D simultaneous localization and mapping (SLAM) is a very impor- tant issue in autonomous robotics. One of the popular algorithms applied as a frontend of SLAM is iterative closest point (ICP). In this paper, the ICP is modelled into a probabilistic framework including both pose estimation and data association steps using expectation maximization (EM). The result derived is that the solu- tion converges to a local minimum if both pose estimation and data association steps employ the same metric. Hence, the measurement model which determines the form of the metric should be the key factor of the algorithm. Then, the point to point, point to plane and plane to plane are analysed in form of their measurement model, which reveals their description of the connection between two scans. Based on analysis, an improvement on point to plane measurement model is presented by estimating the covariance of each plane to relax the model assumption of ICP using eigenvalue decomposition, hence achieving a better solution. The following experiments show a satisfactory performance of the proposed algorithm, in agreement with the theoretic results.