Image-based modeling of built environment from an unmanned aerial system

Highlights

• Automatic image-based modeling for as-built complex constructions

• Application of unmanned aerial system to as-built metric models

• Low-cost three-dimensional as-built metric models of an electrical substation

• Accuracy assessment of the UAS dense surface reconstruction against laser scanner

• Non-invasive three-dimensional reconstruction of complex scenarios

Abstract

In recent years, the modeling from images, called image-based modeling, is emerging as a powerful alternative for the as-built three-dimensional (3D) reconstruction of complex scenarios compared with the use of laser scanner technology or classical topographic methods. Image-based modeling is a non-invasive and low-cost technology that allows 3D reconstruction of objects and scenarios by using only images. However, although several tools currently exist for converting from two-dimensional (2D) to 3D, the specific requirements of quality and completeness for these types of scenarios are difficult to determine.

In this paper, a methodology for automatically reconstructing 3D complex scenarios, particularly electrical substations, using images acquired from an unmanned aerial system is analyzed and proposed. This methodology has been incorporated by the Iberdrola Company as an optimal technique for obtaining low-cost 3D as-built metric models of electrical substations. The results obtained show that this camera-based low-cost system is a competitive alternative to laser scanning systems for modeling as-built complex constructions and can become an essential tool for facility planning and management tasks.

Introduction

Three-dimensional (3D) models have become an essential tool for experts in various fields that provide quality representations of as-built sites, reducing discrepancies between the design and construction phases. These models are used in disciplines including urban and environmental planning [1], [2], [3], cultural heritage documentation [4], [5], [6], building and infrastructure inspection [7], [8], industrial measurement and reverse engineering applications [9], [10], [11] and in the film industry, video game industry and virtual reality applications [12], [13]. However, depending on the application, the specific requirements for a 3D model can be different. In particular, as-built models for engineering and architecture demand metric requirements in terms of accuracy, reliability and completeness, whereas 3D models for the entertainment industry require visual quality in terms of texture-mapping, file size, computational cost, level of detail and ease of use. Therefore, trying to enclose all of these requirements under the same 3D model continues to be difficult for the international scientific community.

To describe complex constructions, laser scanner technology [14] and binomial photogrammetry-computer vision [15], [16] as active and passive techniques, respectively, provide exhaustive and non-invasive 3D documentation methods and can help generate 3D models.

The use of laser scanning technologies for 3D data capture in industrial sites has grown considerably over the last decade over traditional methods for acquiring as-built information, which consists of manual measurements by metric tape and topography [17], [18], [19], due to the improvement in their competitiveness. This competitiveness is due to the rapid increase in the speed and accuracy of the laser scanners in the last decade, while their costs and sizes have been shrinking [20]. However, laser scanner technology requires specialized personnel for the acquisition and processing phases as well as prolonged times for data acquisition and for the processing of the different point clouds.

Conversely, photogrammetric reconstruction, whose use has become popular in recent years due to its hybridization with computer vision [21], has strengthened so-called image-based modeling. The revolution of image-based modeling has been favored because of three main factors: (i) the development of wide-spread computer skills and hence an improvement in the calculation algorithms and possibilities for automation; (ii) flexibility in the camera specifications being allowed to take images, calibrated or not; and (iii) the quality in the results, releasing accurate and reliable as-built 3D models. For example, the computer vision approach for 3D visualization has been applied to document as-planned 3D models for construction sites and to monitor their progress during construction [22], [23]. Although currently there are several image-based modeling tools, none of them guarantee these three results, especially when complex constructions such as electrical substations are considered. In particular, several web-based tools (i.e., Photosynth, Autodesk 123 Catch and Photofly) have been developed for 3D modeling using only images; however, they do not guarantee enough quality in terms of accuracy and reliability [24]. Likewise, other standalone executable tools (i.e., Bundler-PMVS2, VisualSFM) have been developed under the well-known computer vision line, creating structure from motion (SfM), but cannot guarantee enough precision and result in unscaled models [24], [25]. In the case of VisualSfM, the software does not include the global coordinates of the ground control points into the orientation process; this requires that the scale for the reconstructed object or its geo-reference to a global frame must be conducted using a Helmert 3D transformation. According to the close-range photogrammetric community, the most popular tool, PhotoModeler™, which guarantees accurate results, requires manual interaction in the orientation and restitution steps. More recently, some open-source (i.e., Apero-Micmac) [26] and commercial (i.e., PhotoScan™) tools have emerged that combine computer vision and photogrammetric capabilities; however, the former requires command-line usage and does not perform well with multiple images that have oblique geometry, and the latter requires a commercial license.

Considering the limitations detailed above, a hybrid methodology supported by photogrammetry and computer vision that guarantees automation (i.e., converts from 2D to 3D automatically), flexibility (i.e., enables the use of calibrated and non-calibrated cameras) and quality (i.e., provides 3D models with higher resolution than 3D laser scanners) for passing from 2D to 3D has been developed. Photogrammetry Workbench (PW) is a multiplatform software with a user-friendly interface that works with terrestrial or aerial images and considers vertical or oblique geometries [27]. The approach employed by PW software improves the current practice of image-based modeling because it integrates computer vision and photogrammetric algorithms into a smart approach that overcomes the issues of complex objects and scenarios. This is accomplished by combining the last generation descriptors in the extraction and matching steps, combining several lens calibration models under the self-calibration process and combining several stereovision and multiple stereovision algorithms.

This paper is presented in the following structure. After this introduction, Section 2 addresses a detailed description of the developed method for automatic aerial image-based modeling; Section 3 shows the experimental results applied to an electrical outdoor substation, where the shape and size of the elements have high complexity and human interaction with the elements to document may involve security risks, economic costs or even both. Finally, the most relevant concluding remarks are outlined in Section 4.