On Minimizing the Probability of Large Errors in Robust Point Cloud Registration

In solving a model fitting problem, the existence of outliers in the set of measurements can have a devastating effect on the solution accuracy. Traditionally, in order to overcome this problem, robust point cloud registration algorithms are composed of transformation hypothesis generation, followed by hypothesis evaluation aimed at selecting the best hypothesized estimate. Hypotheses evaluation is commonly performed using the sample consensus criterion. However, since this criterion accounts only for the consensus size, it fails when the maximal sample consensus is incorrect. We propose a new hypothesis evaluation approach, generalizing the sample consensus approach, where instead of the sample consensus, the transformation that maximizes the point clouds feature correlation is selected as the best hypothesis. The feature vector at each point contains information such as on local geometry and semantic context. Utilizing this information in the hypotheses evaluation and selection procedure allows for a correct decision even when the hypothesis yielding the maximal sample consensus is false. Consequently, the probability of selecting the correct model increases. We show both mathematically and empirically that substituting the sample consensus criterion with the proposed point cloud feature correlation hypothesis test (PC-FCHT) lowers the probability of large registration errors, compared to using the special case of sample consensus. The proposed PC-FCHT is applicable to any algorithm that follows the hypothesis generation and evaluation scheme, potentially improving the success rates of a wide family of point cloud registration algorithms.