Handling Lens-distortion Removal and Uncalibrated Rectification for a Low-Cost Stereo Vision-Based Navigation System

The architecture of most stereo vision aided navigation systems is based on tracking scene points between stereo images and across time steps. Inexpensive lens systems distort image significantly at its periphery, which requires special pre-processing treatment. Stereo rectification is another necessary pre-processing step to do a faster correspondence matching. In this work, the problems of distortion removal and stereo rectification for a low-cost built in-house stereo vision is studied and a proper solution is proposed that properly handles these tasks. Particularly, the paper demonstrates the poor performance of classical stereo camera calibration for our case with high order of lens distortions, and the successful application of fish-eye distortion model. Moreover, a previously reported method by Fusiello for uncalibrated rectification is extended in this work to cope with a more general case with unequal focal lengths in right and left cameras, as well as in horizontal and vertical directions. The real images captured by a built in-house stereo camera are used to demonstrate the issues and to assess the proposed solution.