以相机作为输入的视觉同时定位与建图(SLAM)系统在地图构建过程中虽然可以保留点云的空间几何信息,但是并没有完全利用环境中物体的语义信息。针对这个问题,对当前主流视觉SLAM系统和基于Faster R-CNN、YOLO等神经网络结构的目标检测算法进行研究。并提出一种有效的点云分割方法,该方法引入支撑平面以提升分割结果的鲁棒性。最后在ORB-SLAM系统的基础上,结合YOLOv3算法进行环境场景的物体检测并保证构建的点云地图具有语义信息。实验结果表明,所提方法可以构建几何信息复杂的语义地图,从而可应用于无人车或机器人的导航工作中。

Visual simultaneous localization and mapping (SLAM) systems that use cameras as input can retain the spatial geometry information of a point cloud in the map construction process. However, such systems do not fully utilize the semantic information of objects in the environment. To address this problem, the mainstream visual SLAM system and object detection algorithms based on neural network structures, such as Faster R-CNN and YOLO, are investigated. Moreover, an effective point cloud segmentation method that adds supporting planes to improve the robustness of the segmentation results is considered. Finally, the YOLOv3 algorithm is combined with ORB-SLAM system to detect objects in the environment scene and ensures that the constructed point cloud map has semantic information. The experimental results demonstrate that the proposed method constructs a semantic map with complex geometric information that can be applied to the navigation of unmanned vehicles or mobile robots.