PREDICTING THE ACCURACY OF PHOTOGRAMMETRIC 3D RECONSTRUCTION FROM CAMERA CALIBRATION PARAMETERS THROUGH A MULTIVARIATE STATISTICAL APPROACH
- 1Dept. of Civil, Environmental, Land, Construction and Chemistry (DICATECh), Politecnico di Bari, Via Orabona 4 - 70125 Bari, Italy
- 2Dept. of Agricultural, Forest and Food Sciences (DISAFA), Università degli Studi di Torino, Largo Braccini 2, 10095, Grugliasco (TO), Italy
Keywords: Unmanned Aerial Vehicles (UAVs), Structure-from-Motion (SfM), Calibration, Principal Components Analysis (PCA), Predictive Analysis
Abstract. Several tools have been introduced to generate accurate 3D models. Among these, Unmanned Aerial Vehicles (UAVs) are an effective low-cost tool to go beyond on-fields effort limits since they allow to fly over areas difficult to reach and to reduce the time needed to collect and process photogrammetric pictures as well. Combining their versatility with Structure from Motion (SfM) techniques efficiency has provided a widely accessible approach to generate accurate photogrammetric products. However, the outcome resolution and coherences also depend on sensor traits. Therefore, UAVs are usually equipped with low-cost non-metric cameras, with the consequent requirement for a calibration procedure to increase the final 3D models accuracy. Although several researchers have highlighted the strong impact of camera calibration parameters on the photogrammetric outcomes, their linkage has not been explored yet. This paper is aimed at investigating their relationship and to propose a novel predicting function of 3D photogrammetric reconstruction accuracy. Such function was estimated thanks to the application of the Principal Components Analysis (PCA) technique. Four photogrammetric UAV flight surveys provided the input data of PCA while an extra dataset was used to validate the results. Once PCA was completed, a synthetic index was proposed and the coefficient of determination was calculated between the index and error components. Synthetic indices values for the various datasets were applied as baseline to detect a predictive function able to assess the northern and eastern error components with a deviation of 0.005 m and 0.003 m, respectively. The proposed approach shows promising and satisfying results for predicting 3D models accuracy.