This paper addresses the efficient state estimation for mixed linear/nonlinear dynamic systems with noisy measurements. Based on a novel density representation - sliced Gaussian mixture density - the decomposition into a (conditionally) linear and nonlinear estimation problem is derived. The systematic approximation procedure minimizing a certain distance measure allows the derivation of (close to) optimal and deterministic estimation results. This leads to high-quality representations of the measurement-conditioned density of the states and, hence, to an overall more efficient estimation process. The performance of the proposed estimator is compared to state-of-the-art estimators, like the well-known marginalized particle filter.