Mean shift algorithm for robust rigid registration between Gaussian Mixture Models

We present a Mean shift (MS) algorithm for solving the rigid point set transformation estimation [1]. Our registration algorithm minimises exactly the Euclidean distance between Gaussian Mixture Models (GMMs). We show experimentally that our algorithm is more robust than previous implementations [1], thanks to both using an annealing framework (to avoid local extrema) and using variable bandwidths in our density estimates. Our approach is applied to 3D real data sets captured with a Lidar scanner and Kinect sensor.