Adding Uncertainty to an Object Detection System for Mobile Robots

摘要

Autonomous mobile robots use vision sensors for navigation due to its ability to provide detailed information of the environment. Visual perception of an environment, is an important capability for mobile robots, so many efforts are leading the research community for such a fundamental and challenging task. An aspect to emphasize is that vision systems have to cope with uncertainty because sensors have noise and the previous knowledge is unclear or inaccurate. In this paper, we propose an uncertainty estimation model applied to an object detection system. The vision system is designed to recognize objects in usual human environments, working on a mobile robot. To calculate the uncertainty, we consider the model accuracy of the system, the probability of detection after the prediction process and the empirical probability of detecting each object according to the distance. The experimental results demonstrate the feasibility and usefulness of incorporating uncertainty information into an object detection system. Finally, the results motivate us to continue improving the uncertainty model in order to use the information generated to strengthen mobile robot navigation systems, as well as for the development of a place categorization system.