Constrained Gaussian Mixture Models Based Scan Matching Method

Abstract

This paper presents a Gaussian mixture model (GMM) based robust scan matching method which implements GMM to represent 2D scan points and improves the accuracy of scan matching. The proposed method transfers each new scan to GMM first, exploiting the covariance of every GMM component to represent scan points. Compared with the conventional GMM based method of scan matching, our technique implements GMM similarity comparison to evaluate the overlaps between scans. In order to get rid of the poor convergence due to the inaccurate initial value given to the iteration process, we proposed a geometryconstraint-based GMM similarity calculation method, which is one contribution of this paper. Another contribution is we propose a dynamic scale factor making the cost function more adapted to different initial value. Experiments on simulated data are employed and the results indicate that our method is able to enlarge the valid range of initial value and accumulate small errors after sequential matchings.