This thesis presents an experimental comparison between active depth sensing with a LiDAR sensor and metric depth estimation from a single image using Metric3D v2. The measurement rig combines a ZED X One 4K monocular camera, an AS-DT1 LiDAR and a Bosch GLM 40 laser distance meter in a fixed geometry. The objective is to determine how three depth outputs—LiDAR, Metric3D ViT-Small and Metric3D ViT-Giant2—dier when estimating the distance to the same visible point and when reconstructing the centre and orientation of a selected planar surface. The point and polygon were selected manually; consequently, the work evaluates distance and surface geometry rather than autonomous object detection.
Measurements were recorded at six nominal conditions from 0.30 m to 2.30 m, with ten repeated acquisitions at each condition, yielding 60 traceable samples. The actual Bosch GLM 40 reading was recorded for each acquisition; in the point test, that reading and the known rig geometry were used to derive the camera-to-red-point reference distance. The LiDAR result was obtained by intersecting the camera ray through the visible red point with a LiDAR-fitted plane, while Metric3D depth was represented by the median value in an 11 ◊ 11 pixel neighbourhood. In the face test, separate planes were fitted to LiDAR and model-derived points within the same manually selected polygon. Bias, mean absolute error, root mean square error, standard deviation, plane-centre dierence, normal-angle dierence and plane-fit RMS were evaluated.
LiDAR produced the lowest overall point mean absolute error of 0.15 m and a pooled within-distance repeatability standard deviation of 0.0020 m. Its error was nevertheless systematically negative, so the result characterises the complete experimental pipeline and must not be interpreted as the intrinsic sensor accuracy. Metric3D ViT-Giant2 achieved lower overall point errors and smaller plane-centre dierences relative to LiDAR than ViT-Small. Conversely, ViT-Small preserved plane orientation more closely: its mean normal-angle dierence relative to the LiDAR plane was 4.79¶, compared with 10.53¶ for ViT-Giant2. A preliminary uncertainty budget included repeatability, reference, rig geometry, pixel selection and the dependence of mean error on distance. The estimated expanded uncertainties at k = 2 were 0.048 m for LiDAR, 0.89 m for ViT-Small and 0.59 m for ViT-Giant2. The nominal 0.30 m condition was retained in the primary analysis; excluding it did not alter the ordering of the three methods by point MAE. The results demonstrate that repeatability, absolute metric accuracy and local geometric fidelity are distinct properties and should be evaluated separately for the intended application.
|
