Lightweight standalone temporal calibration for visual-inertial odometry on resource-constrained embedded systems: a closed-form estimator with statistical hardening
MEASUREMENT: JOURNAL OF THE INTERNATIONAL MEASUREMENT CONFEDERATION, cilt.290, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 290
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.measurement.2026.122778
- Dergi Adı: MEASUREMENT: JOURNAL OF THE INTERNATIONAL MEASUREMENT CONFEDERATION
- Derginin Tarandığı İndeksler: Academic Search Ultimate (EBSCO), Engineering Source (EBSCO), Scopus, Materials Science & Engineering Collection (ProQuest), Technology Collection (ProQuest), Science Citation Index Expanded (SCI-EXPANDED), Compendex, INSPEC
- Karadeniz Teknik Üniversitesi Adresli: Evet
Özet
Accurate temporal calibration between camera and inertial measurement unit (IMU) sensors is a prerequisite for reliable visual-inertial odometry (VIO), yet existing online methods are architecturally inseparable from heavyweight nonlinear optimisers or extended Kalman filter (EKF) back-ends that are impractical for resource- constrained embedded platforms. This paper proposes a standalone, lightweight temporal calibration pipeline that operates independently of any VIO back-end. The pipeline combines: (i) a closed-form scalar projection estimator, derived from first principles at O (1) cost per camera frame and requiring only vector dot products with no iterative solver; (ii) a zero-velocity guard that skips degenerate near-zero-velocity frames, preventing numerical singularity; and (iii) a sliding-window median filter providing outlier-robust offset tracking with a 50% statistical breakdown point. We present the first systematic analysis of estimator behaviour under all four degenerate motion classes of Yang et al. and a comprehensive window-size sensitivity analysis. Critically, we validate the method on all eleven sequences of the EuRoC MAV benchmark with a 35 ms injected synthetic offset, obtaining mean absolute errors of 0.25–0.81 ms across sequences—accuracy competitive with full VIO back-end methods—at only 0.08 ms/frame CPU cost. A dedicated re-excitation experiment further shows that the pipeline recovers from hover intervals of up to 120 s within 8–13 camera frames. Monte Carlo simulation yields 2.18 ms MAE. The proposed pipeline achieves a 98% error reduction over the cross-correlation baseline at 15× lower computational cost, constituting a practical non-iterative solution for edge-computing platforms.