3D point cloud downsampling for 2D indoor scene modelling in mobile robotics

Abstract

Sensory perception and environment modelling are important for autonomous navigation in mobile robotics. 2D discrete grid representations such as the classic 2D occupancy grid maps are a widely used technique in scene representation because of the inherent simplicity and compact representation. In recent years, many 2.5D and 3D grid-based methods have been proposed however, as for the 2D case, a compromise between keeping a low computational bound and reliable sensor interpretation must be kept in order to perform real-world tasks. Assuming the input data in the form of a 3D point-cloud, in this paper we propose a 2D scene modelling approach which converts the 3D data to a 2.5D representation and then to a 2D grid map in an efficient and meaningful manner. The proposed approach incorporates a new rapidly exploring random tree inspired ground-plane detection (RRT-GPD), and an inverse sensor model (ISM) to correctly map 3D to 2.5D and then to 2D grid cells. Experiments were conducted in indoor scenarios with a robotic walker platform equipped with a Microsoft's Kinect One and a LeddarTech's Leddar IS16 sensor. Reported results show an improvement on the representation of non-trivial obstacles (stairs, floor outlets) over a ROS package solution, when applied to a 3D point cloud input.