Evolving visual sonar: Depth from monocular images

摘要

To recover depth from images, the human visual system uses many monocular depth cues, which vision research has only begun to explore. Because a given image can have many possible interpretations, constraints are needed to eliminate ambiguity, and the most powerful constraints are domain specific. As an experiment in the automatic discovery and exploitation of constraints, genetic programming was used to find algorithms for obstacle detection. The algorithms are designed to be a replacement for sonar, returning the location of the nearest obstacle in a given direction. The evolved algorithms worked surprisingly well. Errors were largely transient. The algorithms generalized to both novel views of the office environment and to unseen obstacles. They were combined with a simple reactive wandering program originally written for sonar. The result exhibited good performance in an office environment, colliding only with obstacles outside the robot's field of view. Time to collision results and failure modes are presented. Code is available for download.