Validation of Stereo Matching for Robot Navigation

This paper presents results from experiments on visual stereo matching for robot navigation. Visual features are stereo paired with respect to their pixel position. Stereo triangulating all paired visual features results in a set of landmarks whereof a subset are true landmarks. Constraining the horizontal disparity limits the amount of spurious matches. The stereo matching is validated by finding which landmarks survives short motions measured with a complementary navigation system, like odometry, thus transferring the stereo matching problem from two to three dimensional space and robot motion is estimated from the landmarks surviving the motion. The results from our experiments show that the spurious matching algorithm for stereo matching validation works and that the system is able to estimate the motion.