基于RGB-D相机数据的SLAM算法

摘要

An algorithm for simultaneous localization and mapping of mobile robot based on RGB-D camera data is proposed,which can generate a dense 3D point cloud model of the indoor environment quickly. Firstly,the homogenized ORB features are extracted in the RGB images,using brute-force matcher approach and the RANSAC algorithm,the optimized matching results are obtained. Then the transformation matrix of the continuous frames is solved by PnP,and optimized by nonlinear optimization method. At the same time,key frames are sent to the back-end through the key frame selection mechanism. We use the pose graph optimization in the back-end,and the global pose is optimized by g2o.Based on the transformation matrix,the point clouds in the current frame are transformed into global coordinate to build the dense 3D point cloud map. The scheme is inexpensive and requires only a depth camera as a sensor,we use TUM dataset in the experiments.It cost about 0.182s for each frame,we can conclude that the algorithm can establish a global consistent 3D map of indoor environment rapidly.%提出了一种基于RGB-D相机数据的机器人同步定位与地图构建(SLAM)算法方案,实现了室内环境三维稠密点云地图的快速构建.首先在RGB图中提取均匀化的ORB特征,采用暴力匹配的方式,并使用随机采样一致性(RANSAC)算法得到优化后的匹配结果,然后用PnP解算连续帧的相机位姿变换关系,并用非线性优化方法优化该位姿变换矩阵,同时,通过关键帧选择机制,挑选关键帧,并将得到的关键帧送入后端.后端采用位姿图优化,使用g2o工具箱优化全局位姿.通过位姿变换矩阵将当前帧的点云图变换到世界坐标系中,实现三维稠密点云地图的构建.本文方案成本低廉,仅需要一个深度相机作为传感器,实验中使用的RGB-D数据为TUM制作的数据集,通过实验可以得出,每帧数据平均处理时间为0.182s,本文算法能快速建立全局一致的三维室内环境地图.