Contents I. Introduction
In robotics, Simultaneous Localization and Mapping (SLAM) refers to the problem of building a map of the environment while estimating the robot's pose within it. Traditionallidar-based approaches rely on low-level geometric features such as edges and surface patches to reconstruct the scene [1], [2]. Finding accurate correspondences between two consecutive scans plays an important role in such approaches. With modern hardware, it is possible to estimate the transformation at a high level of detail and scale, and much interest is now turning to semantic labeling this geometry in terms of local regions or predefined objects. The inclusion of rich semantic information enables a much greater range of functionality than geometry alone. For instance, it can help to find better correspondences between two scans by removing unreliable feature points on the moving objects (e.g. pedestrian and vehicle) [3] or by restricting corresponding points to sharing the same semantic label [4]. The dense semantic map of the environment also benefits other modules in a robotic system, such as path planning and human interaction. Thus, this paper focuses on semantic information extraction and its use in a lidar odometry and mapping system.
