Building a Traversability Map on GPU Using RGB-D Camera and Lidar Data

In this article, the author describes the process of optimizing a navigation system for robotic platforms. The focus is on transitioning from a standard 10 cm voxel map to a high-precision 1 mm resolution HD elevation map implemented using CUDA. Leveraging the GPU allowed for significant acceleration in processing data from RGB-D cameras and lidars. During the project, the author encountered two key technical challenges: high depth noise inherent in sensors and vertical odometry instability. The material details data filtering methods and algorithmic approaches that helped minimize errors and achieve stable system performance in real-world conditions. This article is useful for robotics engineers working on localization, mapping, and computer vision.