Abstract
For a mobile robot to navigate safely and efficiently in an outdoor environment, it has to recognize its surrounding terrain. Our robot is equipped with a low-resolution 3D LIDAR and a color camera. The data from both sensors are fused to classify the terrain in front of the robot. Therefore, the ground plane is divided into a grid and each cell is classified as either asphalt, cobblestones, grass or gravel. We use height and intensity features for the LIDAR data and Local ternary patterns for the image data. By additionally taking into account the context-sensitive nature of the terrain, the results can be improved significantly. We present a method based on Conditional Random Fields and compare it with a Markov Random Field based approach. We show that the Conditional Random Field model is better suited for our task. We achieve an average true positive rate of 94.2% for classifying the grid cells into the four terrain classes.