Improved Autonomous Navigation Through Optimal Sensor Outliers

One of the greatest challenges associated with pose estimation using cameras or sensors such as LIDAR is removal of sensor outliers.nbsp; Outliers exist for numerous reasons, such as incorrectly associated feature points, extraneous data points, etc. nbsp;Without proper removal of outliers, an incorrect pose estimate may result, which can lead to difficulty in achieving mission goals for autonomous operations.nbsp; A common approach to remove outliers is the RANSAC algorithm.nbsp; But RANSAC may not be optimal, especially if an incorrect model is used, thus leading to outliers that may pass the RANSAC test but may degrade the pose estimate. nbsp; In Phase I an optimal closed-form approach to replace the RANSAC algorithm was shown that it is based on using the statistical properties of the sensor error in its derivation.nbsp; The heart of the solution is based on an optimally derived pose estimation solution from a Total Least Squares (TLS) approach.nbsp; A byproduct of the TLS solution is the covariance of the sensor residuals, which is the key to remove outliers.nbsp; The Phase I work showed that because the covariance is optimal, it results in a more robust approach to remove outliers than the standard RANSAC algorithm and its variants. nbsp;The Phase I work focused on studying the effectiveness of the newly derived closed-form covariance in a new RANSAC-type algorithm, called the Statistical Optimal RANSAC (SO-RANSAC) algorithm. nbsp; For the Phase II work, SO-RANSAC will be further refined to include the two-camera scenario.nbsp; Invariant methods will also be employed to initialize the SO-RANSAC algorithm. nbsp;Also, a colored-noise filter will be implemented to handle heavy-tail errors associated with feature mis-associations.nbsp; Furthermore, studies will be performed to provide the computational effort required to implement SO-RANSAC on a real spacecraft processor.nbsp; All these efforts will significantly increase the TRL capability of SO-RANSAC, leading to an actual implementation capability.