基于单目SLAM稀疏特征的避障方法

Sparse feature of monocular SLAM based obstacle avoidance

为在使用单目相机的低成本移动机器人上实现避障,提出一种基于现有单目SLAM系统生成的稀疏特征的可通行区域检测方法。通过使用单目SLAM生成的稀疏地标点标识障碍物,根据环境信息恢复构建三维体素地图,生成二维栅格代价地图。为解决单目SLAM中尺度不明确的问题,提出一种基于视觉-轮式编码器的尺度求解器。通过直接使用机器人已有的SLAM框架生成的地标点,在降低系统整合成本同时降低计算量。实验结果表明,在KITTI和DRE数据集上构图的耗时在10 ms~20 ms,可以很好满足在性能受限的设备上的实时避障需求。

To achieve obstacle avoidance on low-cost mobile robots with monocular camera in 3D space,a traversable area detection method based on existing monocular SLAM generated sparse features was proposed.By using sparse landmark points gene-rated through monocular SLAM to mark obstacles,a 3D voxel map was constructed based on environmental information recovery,and a 2D grid cost map was generated.To solve the problem of scale inconsistency problem in monocular SLAM,a scale solver based on wheel encoder and visual odometry was proposed.By directly using the landmark points generated through the existing SLAM framework of the robot,there was no need to compute visual feature points twice.Experiment results on the KITTI and DRE datasets show that the proposed method can construct the map in 10 ms-20 ms,indicating its practicality for real-time obstacle avoidance in devices with constrained resources.