[PDF] ENRICHING 3D BUILDING MODELS BY FAC¸ ADE ELEMENTS

View - Download ENRICHING 3D BUILDING MODELS BY FAC¸ ADE ELEMENTS

Previous PDF

Next PDF

ENRICHING 3D BUILDING MODELS BY FAC¸ADE ELEMENTS BASED ON POINT

CLOUDS AND CONFIDENCE INTERVALS

O. Wysocki, U. Stilla

Photogrammetry and Remote Sensing, TUM School of Engineering and Design, Technical University of Munich (TUM), Munich, Germany

(olaf.wysocki, stilla)@tum.de

KEY WORDS:3D reconstruction, MLS Point clouds, Semantic 3D building models, Confidence interval, Fac¸ade segmentation

1. INTRODUCTION

Semantic 3D building models are used in numerous applications, arewidelyavailable, andcanbeautomaticallyreconstructedupto a level of detail (LoD)2

1. While LoD2 building models display

detailed roof geometries, they lack fac¸ade elements such as win- dows, doors, and underpasses: these building"s features become pivotal for various applications (Wysocki et al., 2021a). Such fac¸ade elements can be acquired by vehicle-mounted mobile laser scanners (MLS) that yield dense, street-level point clouds, which, however, do not provide any semantic information. We propose thus a strategy to refine existing semantic 3D build- ing models using MLS point clouds by modeling absent fac¸ade elements. It is an alternative strategy to from-scratch reconstruc- tion, as we utilize both rich semantic information of building models and detailed geometry of point clouds.

2. RELATED WORK

A method of improving the fac¸ade"s surface accuracy is pro- semantic information. The fac¸ade is refined based on a user- demanded fidelity, under the condition of adequate coverage by MLS observations. Point clouds are segmented by estimating proximity to vertical- and horizontal-like objects extracted from a given city model. The method is, however, limited to error-free- registered semantic 3D models and point clouds (Wysocki et al.,

2021a).

The co-registration uncertainty is addressed by confidence inter- val while fusing point clouds and 3D building models. In the matching step, fine co-registration of point cloud to model is conducted. Then, the fusion is performed using a Bayesian net- work that identifies deviations between the point cloud and the

3D model. In this approach, we assume that MLS sensor origin is

unknown, which limits the identification of such fac¸ade features as windows, doors, and underpasses (Wysocki et al., 2021b). This issue is tackled by a visibility analysis strategy of refining buildings by modeling underpasses. Ray casting of MLS points is conducted on a 3D octree grid from the sensor"s position to the reflection point. Then, a 3D building model is inserted into the grid to identify whether the model"s geometry is conflicted, confirmed, or unknown by MLS observations. Conflicts caused by unmodeled underpasses are identified based on Bayesian reas- oning and road vector features intersecting with fac¸ades. 3D- reconstructed underpasses refine the final, semantic 3D building model (Wysocki et al., 2022a). Detecting and reconstructing fac¸ade features necessities a re- spective validation dataset. To this end, we introduce the TUM-1 https://github.com/OloOcki/awesome-citygmlFAC¸ADE benchmark dataset

2consisting of point clouds an-

notated with 17 fac¸ade-related classes. Furthermore, we review available point cloud datasets and propose a method of extending them for fac¸ade segmentation validation (Wysocki et al., 2022b).

3. FUTURE WORK

In the following project phase, we plan to utilize a visibility ana- lysis strategy as a basis for further building refinements. Besides underpasses, outstanding conflicts on fac¸ades can be caused by openings such as windows and doors. We currently conduct ex- periments to find and reconstruct these openings: The approach combines identified, conflicted areas from visibility analysis and semantic information extracted by deep neural networks trained on the TUM-FAC¸ADE dataset 2. However, 3D model deviations can be caused not only by open- ings: fac¸ade"s plane surface includes neither intruded (e.g., log- gias and arcades) nor extruded (e.g., balconies and moldings) ele- ments. This fact implies that the model"s fac¸ade should consist of more than one given plane. Moreover, each fac¸ade"s re-modeling can impact the correctness of adjacent fac¸ades. As such, these re- finements shall be treated collectively. We investigate these chal- lenges, and we will address them in future works.

ACKNOWLEDGMENTS

This work is supported by the Bavarian State Ministry for Economic Affairs, Regional Development and Energy within the framework of the IuK Bayern projectMoFa3D - Mobile Erfassung von Fassaden mittels 3D Punktwolken, Grant No. IUK643/001. Moreover, the work is conducted within the frame- work of the Leonhard Obermeyer Center at the Technical Univer- sity of Munich (TUM).

REFERENCES

Wysocki, O., Hoegner, L. and Stilla, U., 2022a. Refinement of se- mantic 3D building models by reconstructing underpasses from MLS point clouds.Manuscript submitted for publication. Wysocki, O., Hoegner, L. and Stilla, U., 2022b. TUM-FAC¸ADE: Re- viewing and enriching point cloud benchmarks for fac¸ade segmentation. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information SciencesXLVI-2/W1-2022, pp. 529-536. Wysocki, O., Schwab, B., Hoegner, L., Kolbe, T. and Stilla, U., 2021a. Plastic surgery for 3D city models: A pipeline for automatic geometry re- Remote Sensing and Spatial Information SciencesV-4, pp. 17-24. Wysocki, O., Xu, Y. and Stilla, U., 2021b. Unlocking point cloud poten- tial: Fusing MLS point clouds with semantic 3D building models while considering uncertainty.ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information SciencesVIII-4/W2, pp. 45-52.2 https://mediatum.ub.tum.de/1636761quotesdbs_dbs6.pdfusesText_11

[PDF] aws architecture

[PDF] awt component in javafx

[PDF] awt components in java geeksforgeeks

[PDF] awt components in java in hindi

[PDF] awt components in java ppt

[PDF] awt components in java program

[PDF] awt components in java tutorial point

[PDF] awt components in javatpoint

[PDF] awt controls in java

[PDF] ay tax airline

[PDF] azure devops command line

[PDF] azure fortigate pricing

[PDF] a^nb^n is not regular

[PDF] baby bar essays

[PDF] baby boom 1950