Identifying free space in a robot bird-eye view

Free space detection based on visual clues is an upcoming approach in robotics. Our working domain is the Virtual Rescue League of the RoboCup. In this domain efficient obstacle avoidance is crucial to find victims under challenging conditions. In this study a machine-learning approach is applied to distinguish the difference in visual appearance of obstacles and free space. Omnidirectional camera images are transformed to bird-eye view, which makes comparison with local occupancy maps possible. Bird-eye view images are automatically labeled using Laser Range information, allowing completely autonomous and continuous learning of accurate color models. Two color-based models are compared; a Histogram Method and a Gaussian Mixture Model. Both methods achieve very good performances, with results in a high precision and recall on a typical map from the Rescue League. The Gaussian Mixture Model achieves the best scores with much less parameters on this map, but is beaten by the Histogram Method on real data collected by our Nomad robot. Additionally, the importance of the right color normalization scheme and model parameters is demonstrated in this study.