Camera orientation, calibration and inverse perspective with uncertainties: A Bayesian method applied to area estimation from diverse photographs
Introduction
A large amount of quantitative physical landscape information can be extracted from terrestrial, aerial and satellite imagery using various photogrammetric techniques (Streilein, 1994, Gruen and Li, 1995, Haala and Brenner, 1999, Küng et al., 2012, Feurer and Vinatier, 2018). Inverse perspective methods (e.g. monoplotting), as reviewed by Criminisi, 2001, Förstner and Wrobel, 2016, aim at extracting referenced spatial data from a single picture. Such methods have been used to extract data from either aerial, satellite or terrestrial imagery (Jordan et al., 2005, Bozzini et al., 2012, Murtiyoso et al., 2014, Produit et al., 2016). Inverse perspective methods are particularly used in the study of Earth surface processes and landscape evolution to produce or update geological and geomorphological map data (Warner et al., 1993, Jauregui et al., 2002, Micheletti et al., 2015, Scapozza et al., 2016) or, among others, in civil engineering and building stability assessment (Murtiyoso et al., 2014). The methods have also found a particular echo in the community of cryospheric sciences, as they allow to reconstruct and monitor the evolution of glaciers over different time scales, ranging from centennial to annual fluctuations (Wiesmann et al., 2012, Piermattei et al., 2015, Čekada et al., 2016), as well as further the understanding of glacier mass balance processes (Chapuis et al., 2010).
Inverse perspective methods also allow to tap into a wealth of quantitative information present in large and diverse databases of images readily accessible from the Internet, such as historical records, aerial surveys, or user-contributed pictures. However, these images are of uneven quality: many were taken without scientific intent, often with low-quality lenses and unknown camera parameters, and are sometimes only available in low resolution. These limitations can introduce significant uncertainties and biases in the information obtained from camera orientation and calibration. Useful quantitative data may still be extracted, but accounting for potential lens distortions and quantifying the uncertainty of the results become important.
In this paper, we present a method to address both goals, applying it to the estimation of the area of landscape features, with the determination of the areas of mountain glaciers in mind. The present technique is a two-step process based on Bayesian inference.
First, the unknown camera parameters, including lens optical distortions, are estimated using a Bayesian formulation of the camera orientation with calibration problem, in the form of a spatial resection problem: the posterior probability density function (PDF) of camera parameters is obtained from matches between 2D points in the image and 3D points in world coordinates, together with any available prior information. Bayesian approaches to camera calibration were presented by several authors, either based on finding a single value of the parameters which maximize a posterior distribution (e.g. Valkenburg, 1998, Zhang, 2000) or using the whole resulting posterior distribution more extensively (e.g. Sundareswara and Schrater, 2005). In our work, we keep the full statistical information contained in the posterior distribution of camera parameters, by generating samples distributed according to the posterior distribution using Markov chain Monte Carlo (MCMC) sampling.
In a second step, the posterior PDF of the feature area is derived from the posterior of camera parameters by solving an inverse perspective problem. The outline of the feature of interest is manually traced on the photograph as a polygon, which is then back projected from the 2D image onto the 3D world using a digital elevation model (DEM). This back projection step accounts for uncertainties on the camera parameters, and attempts to model possible systematic errors introduced by the polygon tracing step, together with uncertainties introduced by the DEM. In particular, we propose a model for DEM errors for which both the root mean square error (RMSE) and spatial autocorrelation scale are locally varying. The resulting PDF of the back projected 3D feature area therefore contains information about most of the uncertainties of the process.
More generally, we attempt to unify the camera orientation with calibration, uncertainty modeling and inverse perspective problems into a statistically consistent framework which can be extended to similar classes of problems and uncertainty models.
We stress that given the diverse nature and sources of our target images and the fact that many were taken using low-quality or unknown equipment, the focus of this paper is more on uncertainty estimation than on very accurate photogrammetric techniques. The reconstructed camera location, for example, cannot be expected to be more accurate than a few meters, given that we will be working with low resolution landscape images, and that our 3D ground control point coordinates will not come from precision geodetic sources.
We start by describing our Bayesian formulation of the camera orientation with calibration problem in Section 2, and our polygon back projection method in Section 3. Section 4 describes details of our implementation, including and the posterior probability density sampling process. In Section 5, we present test problems for validation, before discussing the results in Section 6 and concluding in Section 7.
Section snippets
Camera orientation and calibration
Extracting metric measurements from digital images requires estimating the parameters of the imaging camera used to take the picture, such as its position and orientation, focal length, and possibly other optical properties. For typical orientation problems, this may be done by matching points with known 3D world coordinates with their corresponding 2D projections in the image under study. Camera orientation and calibration then consists in finding the camera parameters that best reproduce the
Inverse perspective and uncertainties
We now discuss the inverse perspective step, in the form of the back projection, which is needed to reconstruct the area of a landscape feature from its outline in the image. In Section 3.1, we first describe the back projection process through which we obtain the area S from the camera calibration results. To account for uncertainties in both the polygon tracing process as well as DEM elevations, we derive in 3.2 the posterior distribution of S, assuming imperfect knowledge of the polygon and
Implementation and MCMC sampling
The first step towards implementation is evaluating the Bayesian posterior on camera parameters . We have by now specified the full prior (Eq. (11) with terms discussed in Section 2.4), and likelihood using the camera model and Eqs. (17), (19). We can therefore evaluate the posterior probability density using (2) for any value of the parameter .
Validation
In this section, we setup validation case studies to demonstrate the technique presented in this paper. We use a combination of photographs, both aerial and terrestrial oblique of different origins and quality, to assess key aspects of our method.
First, we consider the problem of camera orientation with calibration, by fitting images for which the camera location is known approximately, and comparing the obtained posterior on the camera position to the available photograph information.
Based on
Discussion
In Section 5.2 we compared the results of the camera calibration procedure to known values of camera parameters. We found that the errors of camera parameters reconstruction for terrestrial oblique pictures are of the same order of magnitude (a few meters) than the basic sources of uncertainty in the method such as the 3D position of the ground control points. This indicates an accurate reconstruction of the camera position by our method, to the intrinsic level of accuracy allowed by the data
Conclusion
In this paper, we presented a novel method for estimating surface area information from landscape features using single aerial and terrestrial photographs. Driven by the goal of characterizing uncertainties on the solutions of inverse perspective problems for archival or non-scientific photographs, we introduced models for errors in input data, as well as for characterizing uncertainties in digital elevation models. We integrated these ingredients into a statistically consistent Bayesian
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
The authors would like to thank the anonymous reviewers for their feedback and comments, which have contributed to significantly improving the manuscript. We also would like to thank the French Institut Gégographique National (IGN) for helpful exchanges, and for providing us with valuable information on the aerial missions. Some map data is copyrighted by OpenStreetMap contributors and available from https://www.openstreetmap.org. Funding: This study is part of the ANR 14-CE03-0006 VIP Mont
References (70)
- et al.
Joining multi-epoch archival aerial images in a single SfM block allows 3-D change detection with almost exclusively image information
ISPRS J. Photogram. Remote Sens.
(2018) - et al.
Road extraction from aerial and satellite images by dynamic programming
ISPRS J. Photogram. Remote Sens.
(1995) - et al.
Extraction of buildings and trees in urban environments
ISPRS J. Photogram. Remote Sens.
(1999) - et al.
Error in a USGS 30-meter digital elevation model and its impact on terrain modeling
J. Hydrol.
(2000) - et al.
A procedure for map updating using digital mono-plotting
Comput. Geosci.
(2002) - et al.
Extracting ecological information from oblique angle terrestrial landscape photographs: performance evaluation of the WSL Monoplotting Tool
Appl. Geogr.
(2015) Towards automation in architectural photogrammetry: CAD-based 3D-feature extraction
ISPRS J. Photogram. Remote Sens.
(1994)- et al.
A new calibration model of camera lens distortion
Pattern Recogn.
(2008) - et al.
A new monoplotting tool to extract georeferenced vector data and orthorectified raster data from oblique non-metric photographs
Int. J. Heritage Digital Era
(2012) Close-range camera calibration
Photogram. Eng.
(1971)
Modelling the spatial distribution of DEM error
Trans. GIS
Monitoring Glacier Changes with the Use of Archive Images: The Example of the Julian Alps (NW Slovenia, NE Italy)
Interpretation of amplitude data from a ground-based radar in combination with terrestrial photogrammetry and visual observations for calving monitoring of Kronebreen, Svalbard
Annals Glaciol.
Accurate Visual Metrology from Single and Multiple Uncalibrated Images, Distinguished Dissertations
Improved modeling of elevation error with geostatistics
GeoInformatica
Objective Bayesian analysis for the Student-t regression model
Biometrika
Emcee: the MCMC Hammer
Publ. Astron. Soc. Pac.
Posterior predictive assessment of model fitness via realized discrepancies
Stat. Sin.
Bayesian Data Analysis
Fundamentals of Computational Photogrammetry
Ensemble samplers with affine invariance
Commun. Appl. Math. Comput. Sci.
Adaptive least squares correlation: a powerful image matching technique
South African J. Photogram., Remote Sens., Cartogr.
Effects of digital elevation model errors on spatially distributed seismic slope stability calculations: an example from Seattle, Washington
Environ. Eng. Geosci.
A global approach for least-squares image matching and surface reconstruction in object space
Photogram. Eng.
Modeling the uncertainty of slope and aspect estimates derived from spatial databases
Geogr. Anal.
Estimation by photogrammetry of the glacier recession on the Cotopaxi Volcano (Ecuador) between 1956 and 1997
Hydrol. Sci. J.
Cited by (7)
Automated point cloud classification using an image-based instance segmentation for structure from motion
2021, Automation in ConstructionCitation Excerpt :We decided to use deep learning because of its outstanding performance over image-related problems [46]. Meanwhile, images from actual locations were selected as an input for PCIS to raise the number of training data because there are currently more public datasets available in the academic society [47], and also numerous images are taken as a part of daily work basis [19,20] due to the common utilization of digital cameras and other capturing tools [48]. Furthermore, other than the point cloud, the parameters such as the focal length, the rotation, and the translation of the capturing tool can also be achieved.
Ice aprons on steep high-alpine slopes: Insights from the Mont-Blanc massif, Western Alps
2023, Journal of GlaciologyVariations in surface area of six ice aprons in the Mont-Blanc massif since the Little Ice Age
2020, Journal of GlaciologyBayesian estimation of glacier surface elevation changes from DEMs
2023, Frontiers in Earth Science