一种融合点线特征的视觉里程计架构设计与定位实现

Design of a Visual Odometry and Localization Based on Point and Line Features Fusing

基于视觉的同时定位与地图构建(SLAM)技术是实现移动机器人自主导航的关键.当机器人处在陌生环境中时,通常会利用周围目标的点特征来估计导航相机的位姿,并利用光束法平差来估计相机位姿和特征空间位置.但如果环境中的特征信息不丰富,则无法准确估计相机轨迹,且欧式坐标与反深度信息下的光束法平差部分条件下不收敛.为此,提出了一种在缺少特征点的环境下通过收集深度相机信息,同时利用点特征与线特征融合的视觉里程计,构建了融合视差角光束法平差与基于线特征的光束法平差的策略,从而使重投影误差达到最小化.最后与其他基于特征的SLAM系统进行比较,实验结果表明,在缺少特征点的真实环境中,系统位姿估计的性能与准确度得到提升.

Point features are mostly extracted in feature based visual simultaneous localization and mapping to estimate camera poses when a robot moves in an unfamiliar environment. However,camera trajectories cannot be estimated accurately if the environment information is not abundant. In this paper,a visual odometry was proposed based on point and line features for RGB-D camera in the environment lacking of feature points. Bundle adjustment (BA) is widely used in estimating camera poses and feature positions. An unavoidable problem of BA with Euclidean coordinates or inverse depth is ill convergence under certain conditions. So a solution was proposed,that integrates parallax bundle adjustment and BA with line features to minimize back-project error. Finally,the proposed approach was compared with other feature based simultaneous localization and mapping (SLAM) system on the dataset TUM. The experiment results show that the proposed approach improves the performance in real scenes lack of point features.