Real Time Obstacle Estimation Based on Dense Stereo Vision for Robotic Lawn Mowers

Abstract

Information on obstacles (e.g., distances, scales, and categories) in the unstructured outdoor environment is crucial for a variety of robotic navigation tasks, such as obstacle avoidance, path planning, and security mechanism for pedestrians. However, most of the previous obstacle estimation methods merely focus on the detection of obstacles. In this paper, we propose a novel automatic obstacle estimation system that not only estimates distance and scale information but also can distinguish pedestrians from other barriers. The designed system comprises two branches, namely obstacle estimation and pedestrian detection. Obstacle estimation infers scales and depth clues from the disparity map using stereo vision, while pedestrian detection detects the pedestrian category by machine learning algorithms. We conduct the experiments on the mowing robot platform, and results show that our proposed system is highly effective. The furthest measurable distance is over 10 meters, while the maximum distance error rate is less than 5%. In addition, the average accuracies of the obstacle estimation and pedestrian detection are 94% and 97.6%, respectively.