SIMPLIFIED BUILDING MODELS EXTRACTION FROM ULTRA-LIGHT UAV IMAGERY

O. Küng, Christoph Strecha, Pascal Fua, Daniel Gurdan, Marius Achtelik, Klaus-Michael Doth, Jan Stumpf2012 · DOI

Abstract Generating detailed simplified building models such as the ones present on Google Earth is often a difficult and lengthy manual task, requiring advanced CAD software and a combination of ground imagery, LIDAR data and blueprints. Nowadays, UAVs such as the AscTec Falcon 8 have reached the maturity to offer an affordable, fast and easy way to capture large amounts of oblique images covering all parts of a building. In this paper we present a state-of-the-art photogrammetry and visual reconstruction pipeline provided by Pix4D applied to medium resolution imagery acquired by such UAVs. The key element of simplified building models extraction is the seamless integration of the outputs of such a pipeline for a final manual refinement step in order to minimize the amount of manual work.

Introduction

Creating accurate 3D building models for virtual globes like Google Earth has traditionally been a labor-intensive process. It typically necessitates a combination of expensive LIDAR data, architectural blueprints, and ground-level imagery. A significant challenge is that ground imagery often misses roof structures, while traditional nadir (top-down) aerial photography fails to capture facades. Furthermore, manual modeling in CAD software is time-consuming. This research addresses these gaps by utilizing affordable, autonomous micro-UAVs to capture high-overlap oblique imagery, automating the core reconstruction pipeline to reduce human effort.

Methodology

The proposed workflow consists of three primary stages: acquisition, automated processing, and CAD integration.

Acquisition

Micro-UAVs, such as the AscTec Falcon 8 octocopter or the senseFly swingletCam, are used to capture a large set of images with 60-80% overlap. The Falcon 8 is particularly effective as it can circle a building, capturing oblique-to-horizontal photos at various altitudes to ensure complete coverage of both roofs and facades.

Circular flight path for oblique imagery acquisition around a target building.

Automated Processing

The automated pipeline leverages state-of-the-art photogrammetry:

  1. Structure From Motion (SfM): The system identifies keypoints across images to recover precise camera positions and orientations.
  2. Dense Matching: Using recovered parameters, a dense 3D point cloud is generated.
  3. Box Model Extraction: A "Level of Detail 1" (LOD 1) box model is automatically generated by projecting 3D points onto a 2D grid to determine the building's footprint and height.

CAD Integration and Refinement

The automated outputs—the LOD 1 model, dense point cloud, and registered images—are imported into Google Sketchup. This allows an operator to efficiently refine the model to LOD 3 (adding architectural details like balconies) by manipulating faces to match the point cloud and applying original photos as textures.

The AscTec Falcon 8 UAV and examples of the oblique imagery captured for reconstruction.

Results and Discussion

The workflow was evaluated on several datasets, including the "Ascending Building" (138 images) and the "EPFL BC Building" (458 images).

Performance and Accuracy

  • Efficiency: Automated processing for 138 images took less than an hour, while smaller sets (36 images) required under 20 minutes.
  • Precision: The mean reprojection error remained consistently low, between 0.6 and 0.9 pixels.
  • Fidelity: Dense matching produced 200,000 to 400,000 points per building, providing a high-quality template for manual refinement.
  • Validation: A model produced via this method was accepted by Google Earth, confirming its professional-grade quality.

Key Insights

A major shift in this approach is the move from minimizing image count to maximizing redundancy. High image overlap makes automatic registration significantly more robust. Furthermore, oblique imagery is identified as the "key element" for reconstructing facades and details under eaves.

Limitations and Future Work

While highly effective, the system faces challenges with repetitive textures (e.g., uniform roof tiles) and reflective surfaces (e.g., glass windows), which can cause noise or holes in the point cloud. Vegetation can also interfere with footprint extraction. Future work will focus on semantic automation—automatically identifying features like balconies—and benchmarking reconstruction quality against official blueprints.

Summary updated

Abstract

Generating detailed simplified building models such as the ones present on Google Earth is often a difficult and lengthy manual task, requiring advanced CAD software and a combination of ground imagery, LIDAR data and blueprints. Nowadays, UAVs such as the AscTec Falcon 8 have reached the maturity to offer an affordable, fast and easy way to capture large amounts of oblique images covering all parts of a building. In this paper we present a state-of-the-art photogrammetry and visual reconstruction pipeline provided by Pix4D applied to medium resolution imagery acquired by such UAVs. The key element of simplified building models extraction is the seamless integration of the outputs of such a pipeline for a final manual refinement step in order to minimize the amount of manual work.
BibTeX
@inproceedings{kung2012simplified,
  author = {O. Küng and Christoph Strecha and Pascal Fua and Daniel Gurdan and Marius Achtelik and Klaus-Michael Doth and Jan Stumpf},
  title = {{SIMPLIFIED BUILDING MODELS EXTRACTION FROM ULTRA-LIGHT UAV IMAGERY}},
  booktitle = {˜The},
  year = {2012},
  doi = {10.5194/isprsarchives-xxxviii-1-c22-217-2011},
}
Figure