Terrain Mapping and Obstacle Detection Using Gaussian Processes

Abstract

In this paper we consider a probabilistic method for extracting terrain maps from a scene and use the information to detect potential navigation obstacles within it. The method uses Gaussian process regression (GPR) to predict an estimate function and its relative uncertainty. To test the new methods, we have arranged two setups: an artificial flat surface with an object in front of the sensors and an outdoor unstructured terrain. Two sensor types have been used to determine the point cloud fed to the system: a 3D laser scanner and a stereo camera pair. The results from both sensor systems show that the estimated maps follow the terrain shape, while protrusions are identified and may be isolated as potential obstacles. Representing the data with a covariance function allows a dramatic reduction of the amount of data to process, while maintaining the statistical properties of the measured and interpolated features.