Autonomous exploration with prediction of the quality of vision-based localization

Abstract

This paper presents an algorithm to perform autonomous exploration with robotic platforms equipped with a stereo-vision system in indoor, unknown and cluttered environments. The accuracy of the vision-based localization depends on the quantity of visible features in the scene captured by the cameras. A Model Predictive Control approach permits to perform the exploration task with obstacle avoidance and taking into account the quality of the scene in order to avoid areas where the visual odometry is likely to fail. Experiments were carried out with a mobile robot to assess the improvement in localization accuracy and coverage for exploration.