Single-lens low-disparity stereo using microlenses

In the framework of the 3-SIS project, we are designing a three-channel miniature camera using an integrated block-based processor and on-chip parallel (SIMD) image processing. This paper describes a single-lens stereo algorithm that we are planning to implement on the narrow-angle channel of the camera. Using microlenses on the sensor delivers a chessboard-like interlaced stereo pair exploiting opposite areas of the lens pupil. As the stereo base is limited by the lens aperture, disparities have very low values; choosing adequate focusing (depending on the scene) and prism deflection angles will give disparities between --1 and 1, which can be measured using local processing only involving directly neighbouring pixels, which is well adapted to the parallel image processing primitives. Subpixel correlation allows real-time, low precision but quasi-dense range information, which may be exploited as an extra clue into the image segmentation process.