Bo Yang

Comparisons of point–line visual–inertial systems can be distorted when trajectories cover different time intervals or when metric estimates receive scale correction. This study asks whether the recorded performance of an EDLines and line-flow system persists under identical temporal support and metric-preserving alignment. The evaluated EDLines (Edge Drawing Lines) visual–inertial navigation system (EDL-VINS) combines conditioned line extraction, pyramidal line-flow association, and Plücker-line factors within a point–line–inertial sliding window. To answer this question, we audit the retained trajectories across all eleven EuRoC sequences, resample the three systems onto a common time grid, and evaluate them without scale correction. The robustness of the resulting comparison is then tested under alternative alignment and timestamp support; gap-aware interpolation; and shorter-horizon, rotational, and archived-output variation analyses. EDL-VINS retains the lowest aggregate translational and rotational trajectory error, although sequence-specific exceptions and short-horizon results show that the ordering is not universal. These findings provide an auditable characterization of the evaluated configuration and show why temporal support, alignment, and data provenance must be explicit in visual–inertial comparisons.

In networked robot localization, remote observations often arrive after the corresponding system state has evolved. Treating these delayed observations as current measurements violates temporal consistency and can degrade estimation accuracy. This study evaluates three delayed-observation processing strategies in a unified finite-window extended Kalman filter (EKF) framework with retrospective updating and forward repropagation: full-time EKF replay, single-threshold latency triggering, and dual-threshold hysteresis triggering. The objective is to determine whether return latency alone can serve as an effective trigger for historical-state correction. The results show that timestamp-consistent updating mitigates the primary errors caused by naive delayed-measurement fusion. Under stochastic latency, however, latency-only triggering produces a distribution-dependent accuracy–workload trade-off, whereas hysteresis improves switching stability primarily near the threshold. These findings clarify the applicability and limitations of latency-based triggering and motivate future trigger designs that jointly account for innovation, uncertainty, observation age, and buffer validity.