본문 바로가기
  • Home

Drone Attitude Estimation under Out-of-Distribution Sensor Noise: A Comparative Study of EKF and LSTM

  • Journal of The Korea Society of Computer and Information
  • Abbr : JKSCI
  • 2026, 31(7), pp.33~43
  • Publisher : The Korean Society Of Computer And Information
  • Research Area : Engineering > Computer Science
  • Received : May 6, 2026
  • Accepted : June 17, 2026
  • Published : July 31, 2026

Bu-Yeon Kim 1 Seok-Won Lee 1

1아주대학교

Accredited

ABSTRACT

The Extended Kalman Filter (EKF) is widely used for drone attitude estimation, but its accuracy can degrade when sensor noise deviates from the assumed noise model. Long Short-Term Memory (LSTM) networks have been applied to IMU-based attitude estimation, yet out-of-distribution (OOD) generalization remains underexplored. This study examines whether an LSTM trained on low-noise data can achieve lower root mean square error (RMSE) than the EKF under unseen high-noise conditions. In the ArduPilot SITL, sensor noise was configured at four levels (L0–L3); a quaternion-output LSTM was trained on L0–L1 and evaluated on L2–L3. The LSTM achieved 21.4% lower mean RMSE than the EKF at L2 and 14.1% lower at L3. However, per-window MAX results indicate that this advantage did not extend to all error metrics. These findings suggest that the LSTM model is better viewed as a complement to the EKF rather than as its replacement under high-noise conditions.

Citation status

* References for papers published after 2025 are currently being built.