FW-NKF: Frequency-Weighted Neural Kalman Filters
Adnan Harun Dogan, Berken Utku Demirel, Christian Holz
AI summary
Problem
Classical Kalman filters assume white Gaussian noise and fail to suppress frequency-dependent disturbances like vibration jitter and bias drift in real-world sensors. Existing deep Kalman variants lack explicit spectral mechanisms to handle these band-limited artifacts.
Approach
The method integrates a learnable causal IIR filter into the Kalman innovation term and trains a neural observation model using a frequency-domain reconstruction loss to selectively attenuate noise-dominated frequencies while preserving state-relevant signals.
Key results
- Up to 10% reduction in state estimation error across four heterogeneous benchmarks
- Marked improvements in orientation accuracy for full-body inertial pose estimation
- Effective suppression of band-limited noise in chaotic systems and real-world MAV datasets
- Ablation studies confirm frequency weighting and deep latent-state modeling jointly drive performance
Why it matters
Provides a robust, frequency-aware alternative for robotics, autonomous navigation, and VR/AR systems that rely on accurate state estimation from noisy IMU and sensor data.
Abstract
Robust state estimation is central to robotic autonomy, yet classical Kalman filters struggle with frequency- dependent disturbances and model mismatch such as sensor vibrations, electromagnetic interference, and periodic noise. Although Deep Kalman Filter (DKF) variants extend the Extended Kalman Filtering (EKF) framework by learning latent transitions, they lack explicit mechanisms to suppress band- limited noise components that typically corrupt sensor mea- surements in real-world scenarios. We introduce the Frequency- Weighted Neural Kalman Filter (FW-NKF), a unified hybrid approach that embeds a causal spectral-shaping operator into the Kalman measurement residual and jointly learns observation, and transition networks. By adapting both the filter spectrum and the latent state representation, FW-NKF attenuates the noise-dominated frequency bands while capturing complex residual structures. We conduct extensive experiments on four heterogeneous benchmarks, including chaotic systems such as multi-dimensional Lorenz systems and full-body inertial pose estimation, and find a reduction in localization error of up to 10% as well as marked improvements in orientation accuracy. Our ablation studies confirm that frequency weighting and deep latent-state modeling contribute to overall performance.