3D modelling of underwater objects using Photogrammetry
Transcription
3D modelling of underwater objects using Photogrammetry
3D modelling of underwater objects using Photogrammetry* Petra Helmholz * Experiments were carried out by Carolyn Martin as part of her master thesis at Curtin University. Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J Motivation Number of Underwater Photogrammetry applications is large, e.g. – Inspection of pipelines and man-made-objects. – Cultural and historical mapping (coral reefs, ship wrecks). Underwater Photogrammetry is often expansive (equipment, process of taken the images) and the image quality is often poor. In the last decade, a number of waterproof consumer cameras e.g. “action cameras” became available (often for less than $1,000 ). Many use fisheye lenses which are often not suitable for geometric accurate photogrammetric applications . 21/11/2013 Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J WALIS Forum 2013 Page 2 Image taken from http://gopro.com/ Image taken from www.ioncamera.com Motivation New structure-from-motion based programs provide “everyone” the possibility to create 3D models based on images (e.g. VisualSFM, 123D Catch, …) or ? No or little prior knowledge is required to use these programs. or ? Aim: – Test of geometric accuracy of a underwater camera with fisheye lens (GoPro Hero 3 – black edition) – Test of Photogrammetric solution against Structure-From-Motion solution 21/11/2013 Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J WALIS Forum 2013 Images taken from http://www.cs.cornell.edu Page 3 Outline Background – Interior orientation parameters and camera calibration – Relative orientation and scale – Structure from motion Experiments – Performance above water (using photogrammetric workflow) – Performance under water (using photogrammetric workflow) – Accuracy comparison of photogrammetric workflow to structure-frommotion approach Conclusions and Outlook 21/11/2013 Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J WALIS Forum 2013 Page 4 Interior orientation parameters Interior orientation defines the camera parameters required for reconstructing the bundle of rays. Parameters are: • The camera’s calibrated focal length (c). • The position of the principle point (PP) in the image plane (x’P, y’P). • Deviations from the ideal sensor model (e.g. distortion Dr’) main distortion component is of radial structure. A photogrammetric analysis is possible ONLY when the interior orientation parameters are known. 21/11/2013 Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J WALIS Forum 2013 Page 5 Sensor plane Projection centre O c Camera body Principle Point and Principle Point Offset The calibrated focal length c: • Is the distance of the projection centre O from the image plane. Pixel coordinate system (column, row) Image coordinate system (x, y, (c)) Principal point: • • y c Point of intersection of the perpendicular from the image perspective centre O with the image plane. yp If the camera is properly adjusted, the principal point is close to the origin of the image coordinate system. xp Principle Point offset (xp, yp): is the offset of the principal point (PP) from the origin of the image coordinate system. 21/11/2013 Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J WALIS Forum 2013 PP Page 6 x Lens distortion Lens distortion Dr’ is present when the actual light ray in the image space (O’ P’) is not parallel to the corresponding light ray in object space (OP). Radial lens distortion: the functional model for radial lens distortion is an odd-powered polynomial as a function of radial distance Dr k1r 3 k2 r 5 k3r 7 The coefficients, k1, k2 and k3 are determined directly (calibration). O 21/11/2013 Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J WALIS Forum 2013 3 2 1 Image plane Dr Image space P’ Deformed Image Surface O Refracted Ray Path (Model and Actual) Object Point r3 r2 r1 c r PP Page 7 Object space P Camera calibration Suitable target field of object points. Test field is imaged from several camera stations, ensuring good ray intersection and filling the image format. Convergent image configuration, rather than the “normal case” traditionally used for aerial photography. At least 8 images are acquired, each imaging as many of the test field targets as possible. The test field is imaged perpendicularly and obliquely and each image set should have one image rotated by 90° around the optical axis. Take consistently sharp images (aperture setting, fast shutter speed, low ISO). 21/11/2013 Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J WALIS Forum 2013 Page 8 test field Camera positions Examples for coded targets (Image taken from http://www.iwitnessphoto.com/ ) Orientation Procedures GCP (X, Y, Z) ≥3 ≥3 Adopted from Luhmann et al, 2006 (p. 202) ≥3 ≥3 Space resection Bundle adjustment Relative orientation nl 1 2 Exterior Orientation ≥2 21/11/2013 WALIS Forum 2013 Absolute orientation nN ≥2 Space intersection Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J ≥5 Image points (x, y) nP 1 New points (X, Y, Z) Page 9 Video Structure From Motion Simultaneously estimating 3D geometry (structure) and camera position (motion) (Ullman, 1979). Solves the triangulation (intersection), two-frame structure from motion (relative orientation) & bundle adjustment using the projective geometry. Approach mainly used in the Computer Vision Building Rome in a Day1 community to process big unorganised and unstructured image datasets. For instance, “Building Rome in a Day” project1 Images taken from Google and Flickr Easy to use and looks good, but what is the geometric accuracy? 1 Agarwal, S., Furukawa, Y. , Snavely, N., Simon, I., Curless, B., Seitz, S.M., Szeliski, R., 2011. Building Rome in a Day. In Communications of the ACM, Vol. 54, No. 10, Pages 105-112, October 2011. 21/11/2013 Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J WALIS Forum 2013 Page 10 Equipment Cameras: Nikon D80 (SLR used in Close Range Photogrammetry) and GoPro Hero black edition (consumer camera with fisheye lens). Calibration field (box) with coded targets (above water). Calibration frame with reflecting Calibration frame (underwater) targets (underwater). Software: iWitnessPro (Photogrammtric), VisualSfM (StructureFrom-Motion). Images taken from Carolyn Martin 21/11/2013 Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J WALIS Forum 2013 Page 11 Calibration box (above water) Camera calibration above water Comparison of the parameter of the interior orientation from the Nikon and GoPro camera using the calibration field with coded targets. 10 images per camera with one rotated image. Scale was introduced using observed distances between reflective targets. Both datasets were transformed in the same local coordinate system defined by 3 reference points. Reflective targets could get pick up automatically with Photogrammetric software in the images taken with the Nikon SLR but not with the GoPro (e.g. motion blur). 21/11/2013 Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J WALIS Forum 2013 Image taken from Carolyn Martin Page 12 Calibration box (above water) Camera calibration above water Nikon D80 GoPro c 17.943 mm 2.755 mm xP -0.016 mm -0.070 mm yP -0.115 mm 0.109 mm k1 5.488∙10-4 4.408∙10-2 k2 -6.109∙10-7 2.166∙10-3 k3 5.188∙10-10 -6.256∙10-5 Quality index 1:2,100 1:1,900 Accuracy of referencing 0.63 pixels 1.77 pixels No. of ref. points 88 82 21/11/2013 Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J WALIS Forum 2013 Radial Distortion Grid Nikon D80: Two orders of a magnitude difference. Radial Distortion Grid GoPro: Page 13 Accuracy check above water Residuals between Nikon (reference) and GoPro. Residuals (scale factor 30) Magnitude of residuals [mm] average 2.19 max 4.05 z Standard deviations [mm] X 1.40 Y 1.35 Z 1.60 21/11/2013 Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J y WALIS Forum 2013 x Page 14 Camera calibration underwater Use of only the GoPro Camera. Reference is given by reference frame. Experiment was performed in a water tank at Curtin University (thanks to the Centre for Marine Science and Technology (CMST)). Lots of motion blur during the images are taken only a small subset of images was useable (6 images). Motion blur, low light and low visibility is in general an issue in Underwater Photogrammetry. 21/11/2013 Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J WALIS Forum 2013 Calibration frame (underwater) Images taken from Carolyn Martin Page 15 Camera calibration underwater GoPro camera Above water Under water c 2.755 mm 3.607 mm xP -0.070 mm -0.030 mm yP 0.109 mm 0.121 mm k1 4.408∙10-2 1.845∙10-2 k2 2.166∙ -1.550∙ k3 -6.256∙10-5 3.747∙10-6 Quality index 1:1,900 1:2,400 Accuracy of referencing 1.77 pixels 0.65 pixels No. of ref. points 82 34 21/11/2013 Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J 10-3 WALIS Forum 2013 10-3 Distortion by fisheye lens is smaller. Similar to Nikon in previous test (0.63 pixels). Smaller dataset. Page 16 Radial Distortion Grid GoPro (above water and below water) Accuracy check underwater Residuals between calibration frame (reference) and GoPro Magnitude of residuals [mm] Residuals (scale factor 30) average 1.36 instead of 2.19 mm max 3.09 instead of 4.05 mm Standard deviations [mm] X 1.24 instead of 1.40 mm Y 0.70 instead of 1.35 mm Z 1.17 instead of 1.60 mm z y 21/11/2013 Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J WALIS Forum 2013 x Page 17 Comparison to the Structure-from-motion app. Software output is a point cloud, an orthoimage, a textured model and a video file. Observations in the images are not possible (no space intersection). Points of interest could be not extracted in the point cloud or the mesh. Extracted 3D Point Cloud Back frame Front frame Ortho-image Data taken from Carolyn Martin. 21/11/2013 Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J WALIS Forum 2013 Page 18 Conclusions Fisheye effect above water has a very strong influence and leads to high radial distortions (e.g. epipolar line is actually a curved line). The fisheye effect decreases by underwater application (distortion pattern and geometric accuracy improves). Structure-from-motion can often achieve good models but struggles with difficult scenes like in this extreme example. Further investigations in the geometric accuracy are necessary. Curved epipolar line 21/11/2013 Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J WALIS Forum 2013 Page 19 Outlook In 2008 HMAS Sydney and HSK Kormoran was found in 2,500 m depth off Geralton’s coastline by the Finding Sydney Foundation. Images were taken without the application of 3D modelling but some 3D models could be extracted from the dataset (Hollick et al., 2013). HSK Kormoran foredeck 2D Photographs © 2008 Australian War Memorial via The Finding Sydney Foundation. 21/11/2013 Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J WALIS Forum 2013 Page 20 Video Outlook Revisit of HMAS Sydney is scheduled for 2014/2015. Equipment for the dive (ROV and cameras) is available. Goal: to create a 3D model of the ship wreck as complete as possible with a medium to good geometric accuracy. Team at Curtin working on existing data set and on the revisit plans (Andrew Hutchison, Andrew Woods, Petra Helmholz). Models will be used in exhibitions and also for further analysis. HSK Kormoran foredeck Image taken from Hollick et al. (2013) 21/11/2013 Curtin University is a trademark of Curtin University of Technology CRICOS Provider Code 00301J WALIS Forum 2013 Page 21