基于半直接视觉里程计的SLAM稠密重建

SLAM with dense reconstruction based on semi-direct visual odometry

针对特征法视觉同步定位与地图构建(simultaneous localization and mapping,SLAM)耗时久、纹理缺失场景易跟踪失败而只能重建稀疏点云的问题,提出一种基于半直接视觉里程计的SLAM稠密重建算法。该算法以ORB-SLAM2框架为基础,视觉里程计采用直接法最小化光度误差求解相机初始位姿,采用特征法最小化重投影误差优化位姿,提高相机位姿的输出频率;以一种冗余关键帧删除算法计算帧间相对运动距离,删除视野重叠的关键帧;使用筛选后的关键帧进行闭环检测,构建稠密点云地图,滤波后转换为存储效率更高的八叉树地图。结果表明,所提算法能有效解决相机快速运动场景和纹理缺失场景跟踪失败的问题,实现三维地图重建,具有较高的定位精度与实时性。

Feature-based visual SLAM(simultaneous localization and mapping)takes a long time and the scenes with missing texture are easy to be tracking failure.Only sparse point clouds can be reconstructed.Therefore a SLAM with dense reconstruction algorithm based on semi-direct visual odometry was proposed and the algorithm was based on the ORB-SLAM2 framework.The visual odometry used the direct method to minimize the photometric error and to solve the initial pose of the camera.It utilized the feature method to minimize the reprojection error to optimize the pose,which increased the output frequency of the camera pose.A redundant key frame deletion algorithm was proposed to calculate the relative movement distance between frames and deleted the key frames with overlapping fields of view.The filtered key frames were applied for closed-loop detection to construct a dense point cloud map,which was converted into an octree map with higher storage efficiency after filtering.Experimental results show that the algorithm can effectively solve the problem of tracking failure in fast camera motion scenes or texture missing scenes.It realizes three-dimensional map reconstruction,and it has high positioning accuracy and real-time performance.