A Newton-Type Filter for Nonlinear Systems with Unknown Noise Distribution

Nonlinear filters have always been the core technology in many fields such as tracking, navigation and localization. However, their performance is often plagued by noise distribution characteristics. This paper proposes a novel Newton-type filter (NTF), in which a measurement-driven mapping is designed to assess the quality of measurements, as a stepsize vector allowing for the adjustment of Newton optimization. Unlike traditional Kalman-based filters, it does not require the computation of covariance or knowledge of noise distributions. Experiments on KITTI data set and indoor robot localization show that the proposed algorithm outperforms traditional filters.