Stereoscopic vision for a humanoid robot using genetic programming

In this paper we introduce a new approach to adaptive stereoscopic vision. We use genetic programming, where the input to the individuals is raw pixel data from stereo image-pairs acquired by two CCD cameras. The output from the individuals is the disparity map, which is transformed to a 3D map of the captured scene using triangulation. The used genetic engine evolves machine-coded individuals, and can thereby reach high performance on weak computer architectures. The evolved individuals have an average disparity-error of 1.5 pixels, which is equivalent to an uncertainty of about 10% of the true distance. This work is motivated by applications to the control of autonomous humanoid robots - The Humanoid Project at Chalmers.