BumpyPatch: Heightmap-based Outdoor Point Cloud Segmentation to Find Less Bumpy Road

Autonomous mobile robots operate in a range of environments, from controlled indoor settings to unpredictable outdoor terrains. These varied conditions present challenges that require advanced navigation systems for their safe and efficient operation. A key component of this navigation is accurately assessing the ground texture. Any misjudgments can jeopardize both the robot's sensitive equipment and its carried cargo. In this research, we propose a novel method that uses heightmaps created by mapping the
$\mathrm{z}$
-coordinates of 3D LiDAR-derived point cloud data to grayscale pixel values for evaluating outdoor ground textures. This approach effectively converts point cloud data, providing information to assist mobile robots in navigating less bumpy roads in outdoor settings. We present classification techniques for terrains based on the environment's nature: static, which pertains to individual point cloud files representing completed scenes, and dynamic, related to the real-time point cloud data captured by moving robots. For both static and dynamic environments, we introduce tailored heightmap classifiers, incorporating Inertial Measurement Unit (IMU) insights to consider the robot motion influenced by terrain texture. Our proposed method demonstrates superior performance compared to existing methods that analyze the point cloud directly and perform texture analysis with high accuracy in both static and dynamic environments. The code can be downloaded from https://github.com/zzziito/BumpyPatch.