Robust State Estimation and Sensor Fusion for UAVs
How sensor physics and failure modes shape fusion design Choosing between EKF, UKF and particle filters for UAVs Designing a state vector and observation models that converge Time synchronization and sensor calibration that prevent divergence Operational robustness: outlier rejection, FDI, and graceful degradation Practical implementation checklist for reliable IMU–GPS–Vision fusion Accurate UAV…
Sensor performance and failure modes dictate how a UAV's fusion system is designed. Designers must understand the physics of MEMS inertial sensors, GPS, and vision systems. MEMS IMUs have noise, bias, scale, axis misalignments, and temperature effects. These map to process noise and bias models in filters. GNSS provides position but has multipath, NLOS, intermittent fixes, carrier slips, and spoofing.
Vision sensors depend on shutter type, latency, exposure, motion blur and low-contrast scenes. Sensors must be described as nominal models, failure modes, and recovery behavior. This allows the fusion system to gate updates and degrade gracefully.
Two main filter choices exist: EKF, UKF and particle filters. EKF's lightweight, deterministic nature makes it suitable for microcontrollers. UKF captures higher-order nonlinearities but is heavier. Particle filters are flexible but computationally expensive. The author recommends error-state, multiplicative quaternion EKF for attitude handling in VIO on resource-constrained UAVs.
This avoids quaternion renormalization issues and keeps linearization small. For higher-rate IMU and lower-rate visual sensors, MSCKF or optimization-based sliding-window solvers are common.
State design is engineering, trading expressiveness for numerical cost and observability. A compact state vector often includes error-state EKF pattern elements: position, velocity, quaternion, IMU and gyro biases, extrinsics and camera-IMU time offset. The error-state representation simplifies linearization and avoids quaternion singularities.
Covariance updates occur in error space, with multiplicative corrections to the quaternion. Observation models must account for sensor physics, failure modes and recovery behavior. This leads to defensible gating and graceful degradation of the fusion system.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
