Achieving Consistent Uncertainty Estimates with RANSAC-Based Algorithms

Clark N. Taylor

Abstract: When using visual processing algorithms to perform navigation, most previous works have focused on finding the best ”maximum likelihood” estimate possible given the measurements. When performing navigation, however, having an accurate estimate of the uncertainty is often as important as the estimate itself. Uncertainty estimates are typically generated by assuming known-covariance Gaussian noise on the measurements and propagating this information through the navigation system. Unfortunately, the measurements from a visual sensor are not like the measurements from a typical sensor where the measurement itself is a direct output of the sensor. Instead, the output of a visual sensor is an image which is processed to find feature locations, which become the measurements for downstream algorithms. These feature location measurements are known to be corrupted with significant outliers, meaning that generally some outlier rejection algorithm is used to reduce the measurements down to just the inliers. Unfortunately, the effects of outlier rejection algorithms on the uncertainty estimates have not been previously studied in great detail. In this paper, we analyze one of the most common outlier rejection algorithms, RANSAC, for its effect on uncertainty. We find that RANSAC causes significant errors (inconsistency) in the uncertainty estimation and attempt to explain this effect. After discussing the causes of this inconsistency, we propose an approach to overcome this error and present results demonstrating the efficacy of this approach.
Published in: Proceedings of the 2016 International Technical Meeting of The Institute of Navigation
January 25 - 28, 2016
Hyatt Regency Monterey
Monterey, California
Pages: 494 - 500
Cite this article: Taylor, Clark N., "Achieving Consistent Uncertainty Estimates with RANSAC-Based Algorithms," Proceedings of the 2016 International Technical Meeting of The Institute of Navigation, Monterey, California, January 2016, pp. 494-500.
Full Paper: ION Members/Non-Members: 1 Download Credit
Sign In