Variables Influencing the Accuracy of 3D Modeling of Existing Roads Using Consumer Cameras in Aerial Photogrammetry.

Point cloud (PC) generation from photogrammetry⁻remotely piloted aircraft systems (RPAS) at high spatial and temporal resolution and accuracy is of increasing importance for many applications. For several years, photogrammetry⁻RPAS has been used to recover civil engineering works such as digital elevation models (DEMs), triangle irregular networks (TINs), contour levels, orthophotographs, etc. This study analyzes the influence of variables involved in the accuracy of PC generation over asphalt shapes and determines the most influential variable based on the development of an artificial neural network (ANN) with patterns identified in the test flights. The input variables were those involved, and output was the three-dimension root mean square error (3D-RMSE) of the PC in each ground control point (GCP). The result of the study shows that the most influential variable over PC accuracy is the modulation transfer function 50 (MTF50). In addition, the study obtained an average 3D-RMSE of 1 cm. The results can be used by the scientific and civil engineering communities to consider MTF50 variables in obtaining images from RPAS cameras and to predict the accuracy of a PC over asphalt based on the ANN developed. Also, this ANN could be the beginning of a large database containing patterns from several cameras and lenses in the world market.

MatchingLand, geospatial data testbed for the assessment of matching methods.

This article presents datasets prepared with the aim of helping the evaluation of geospatial matching methods for vector data. These datasets were built up from mapping data produced by official Spanish mapping agencies. The testbed supplied encompasses the three geometry types: point, line and area. Initial datasets were submitted to geometric transformations in order to generate synthetic datasets. These transformations represent factors that might influence the performance of geospatial matching methods, like the morphology of linear or areal features, systematic transformations, and random disturbance over initial data. We call our 11 GiB benchmark data 'MatchingLand' and we hope it can be useful for the geographic information science research community.