Progressive Gaussian Mixture Reduction

For estimation and fusion tasks it is inevitable to approximate a Gaussian mixture by one with fewer components to keep the complexity bounded. Appropriate approximations can be typically generated by exploiting the redundancy in the shape description of the original mixture. In contrast to the common approach of successively merging pairs of components to maintain a desired complexity, the novel Gaussian mixture reduction algorithm introduced in this paper avoids to directly reduce the original Gaussian mixture. Instead, an approximate mixture is generated from scratch by employing homotopy continuation. This allows starting the approximation with a single Gaussian, which is constantly adapted to the progressively incorporated true Gaussian mixture. Whenever a user-defined bound on the deviation of the approximation cannot be maintained during the continuation, further components are added to the approximation. This facilitates significantly reducing the number of components even for complex Gaussian mixtures.