HIET: Height Images for Enhanced Terrain Perception for Robot Navigation

The representation of unknown environment poses a significant challenge in the field of robotics. The existing methods leverage elevation maps constructed from point cloud data from sensors like 3D-LiDAR or Stereo camera. Robots utilizes this map for navigating through environments and achieving perceptual locomotion control. In this paper, we introduce a grid based elevation mapping pipeline which harnesses the height information from 3D point cloud to generate a 2D height image as intermediate format and use to create 2.5D elevation map. The proposed height image based approach is a new way of getting elevation map and is equally holds relevance due to its ability to process lower-dimensional data while still constructing a 2.5D map. To achieve this, we first find the receptive field of image by establishing the correspondence between image pixels and 3D points through projection. In the receptive field of the image, the pixel intensities are computed by checking the average height of points resulting in an image with height intensities. Further, this height image is post-processed as a gridmap layer to have structured and informative representation. Our pipeline is specifically engineered to achieve good processing speed and it is highly configurable to allow user to configure the map parameters and enable a wide variety of environment representation. We carried out extensive testing on four open source data namely LioSAM, Rellis3D and Fast_LIO. We also showcase its practical applicability by deploying across multiple robotic platforms with diverse configurations. The proposed pipeline includes a ROS2 interface for easy integration with any Autonomous Navigation System (ANS) stack.