Probabilistic outlier removal for robust landmark identification in stereo vision based SLAM

We consider the problem of performing simultaneous localisation and mapping (SLAM) with a stereo vision sensor, where image features are matched and triangulated for use as 3D landmarks. We explain how we obtain 3D landmark measurements and derive a Gaussian noise model for these measurements. We then argue that the classic way of removing outliers from stereo image features, by estimating fundamental matrices, has limitations. We propose instead the use of a probabilistic measure in determining consensus sets of hypotheses generated in a RANSAC framework. In order to test the performance of this approach we incorporate it into an EKF SLAM system, which is notorious for its sensitivity to landmark mismatches. We measure the accuracy achieved on outdoor datasets, using DGPS as ground truth, and compare it to two other standard SLAM algorithms. We find that the proposed system outperforms the others significantly.