基于概率选取随机特征点的单目视觉SLAM 方法

同时定位与地图构建（SLAM: Simultaneous Localization And Mapping）是自主机器人导航研究的重要内容，基于扩展卡尔曼滤波器(EKF: Extended Kalman filter)的单目视觉SLAM方法是解决该问题的一种重要手段。但由于计算复杂度过高的原因，它对特征点数量及稳定性有严格要求，传统的方法是选取少量窗格型显著角点或直线段作为特征点，限制了该方法的应用范围。本文提出基于随机概率统计的SIFT（Scale-invariant feature transform）特征点选取方法，在保证特征点分布相对均匀的前提下，有效的控制了特征点的总体数量，从而达到了减少单目视觉EKF-SLAM方法应用限制的目的。实验研究表明，该特征点选取方法对不同场景均具有较高的稳定性，并且一定程度上提高了算法的收敛速度。

Visual Monocular SLAM Based on Probabilistic Selection of Random Feature Points

For the problem of visual monocular robot SLAM (simultaneous localization and mapping), a random feature point selection method based on probability and statistics of SIFT (scale-invariant feature transform) feature points is proposed by using SIFT feature points and the inverse depth method. On the assumption of relative uniform distribution of the feature points, the total amount of feature points is constrained effectively, and the application restriction of the visual monocular EKF-SLAM (extended Kalman filtering SLAM) is relaxed. Experiments show that this feature point selection method is of high stability in different scenes, and improves the convergence speed to some extent.