Camera-LiDAR-based Indirect Evaluation of Estimated Driving Trajectories for Aircraft MRO Inspection Robots

Abstract

Evaluating a robot’s estimated trajectory typically requires a ground truth trajectory, but establishing the necessary measurement infrastructure in large aircraft hangars is costly and operationally burdensome. Although alternative methods reduce this burden, they often impose other requirements in the environment, such as closed-loop routes, marker installation, or specific environmental geometry, limiting their applicability to aircraft maintenance, repair, and overhaul (MRO) inspection. This study proposes an indirect evaluation method for aircraft inspection robots without requiring a ground truth trajectory or modifications to the inspection environment. The proposed method detects relative spatial inconsistencies between the estimated trajectory and single-camera observations. It uses an object detection-tracking algorithm and the aircraft fuselage as a reference structure to evaluate the estimated trajectory by comparing it with camera observations in the image plane. To support reliable tracking without additional training, we construct herein a visual foundation model (VFM)-based detection-tracking pipeline with preprocessing and postprocessing. Field experiments in an actual aircraft exterior inspection environment compare the evaluation results of the original estimated trajectories with those obtained from drift-injected trajectories. The original trajectories achieve a mean trajectory-alignment F1 score of 0.9195, while alignment decreases as the trajectory portions subject to drift expand. These results support the method’s ability to detect spatial inconsistencies between estimated trajectories and camera observations without ground truth trajectories.