Prediction of the scene quality for stereo vision-based autonomous navigation

This paper presents an autonomous navigation architecture for a robot using stereo vision-based localisation. The main contribution is the prediction of the quality of future localisation of the system in order to detect and avoid areas where vision-based localisation may fail, due to lack of texture in the scene. A criterion based on the estimation of future visible landmarks, considering uncertainties on landmarks and camera positions, is integrated in a Model Predictive Control loop to compute safe trajectories with respect to the visual localisation. The system was tested on a mobile robot and the obtained results demonstrate the effectiveness of our method.