Biologically Motivated Qualitative Landmark Recognition Using Visual Cues

An autonomous robotic vehicle which needs to move around in an unstructured environment should be able to detect obstacles in its patha and should know its own position at any given time with respect to a global map. Thie requires that the robot detect and recognize certain landmarks and infer its position in the map from the location of the landmarks. Reliable recognition of a predefined set of generic landmarks using a visual sensor is a difficult problem. The difficulties are posed by the unstructured, cluttered environment. Changes in viewing angle, occlusions, sensor noise present further problems. The current work in lanmark recognition concentrates mostly on approaches to efficient navigation providing the robot with some form of detailed representation of the environment. We propose a biologically-based artificial intelligence approach utilizing visual cues suitable for both indoor and outdoor landmark recognition. We first apply a biologically motivated filter to extract edge and boundry features from visual images. We then apply qualitative techniques like perpetual grouping to hypothesize and verify man-made structures. The natural textures such as grass and water in the outdoors are segmented and recognized using texture boundaries obtained using the output of the same filter. Together these techniques provide a reliable recognition of indoor as well as outdoor landmarks.