Photogrammetric Performance of an Ultra Light Weight Swinglet UAV

Transcription

1 Photogrammetric Performance of an Ultra Light Weight Swinglet UAV J. Vallet, F. Panissod, C. Strecha, M. Tracol UAV-g Unmanned Aerial Vehicle in Geomatics September 14-16, 2011ETH Zurich

2 Summary Introduction Vehicule and Sensors description Methodology Experiences and datasets Analysis and results Conclusion&outlooks

3 Introduction Development of last decade: Lighter sensors for georeferencing: GPS, attitude, MEMS High resolution compact digital camera New image correlation algorithms Powerful standard computers UAV auto-drived for non specialist Widening of digital photogrammetry application field Low cost Photogrammetry for non specialist Applications with UAV

4 Introduction Case study of ultralight UAV Low cost camera Low cost sensors for Georeference Study mapping performance of such system Study workflow performance: standard photogrammetric software new generation

5 Vehicule Ultra light vehicule: Sensefly Swinglet less than 500g u-blox GPS chip + MEMS attitude Payload of 125g Autopilot to drive on flight line and auto image trigger ~25 min autonomy Flight restriction: wind (< 20km/h), keep it onsight

6 Image sensor Canon Ixus-120IS 12 Megapixels Pixel size: 1.54 um Chip size: 6.16x4.62mm Autofocus, no fixed focal length

7 Georeference C/A code GPS chips Position + velocity recording at 1Hz Attitude recording with accuracy of <5 No possibility to process differential C/A trajectory accuracy < 5-10m Information tagged to EXIF data of each image

8 Processing Classical photogrammetric Software: SocetSet + Bingo-F Automated new algorithm: Pix4D Automatic Tie point matching, auto camera calibration DSM extraction, orthomosaic Comparison with reference data GCP s LiDAR/photogrammetric reference dataset

9 Analysis methodology Camera calibration: focal, PPA, dist. (radial/tangential) AT results (bundle adjustment): residuals DTM/DSM comparison 25cm Grid comparison altimetric accuracy GCP Noise analysis Ortho accuracy: GCP feature digitizing

10 Experiences Swinglet flight 72 images, GSD ~5cm 8 flight lines irregular overlap (wind) Helimap System flight same pattern LiDAR ~6 pt/m 2 /line 35 images, GSD 4cm Ground measurements 12 GCP s (Static DGPS)

11 Workflow Swinglet Images Swinglet Trajectory GCP Pix4D Tie/GCP points BA Pix4D BA Bingo-F Image orientation Camera parameters Image orientation Camera parameters SocetSet NGATE Pix4D image corr. DTM socet SocetSet mosaic Pix4D mosaic DSM pix4d Ortho socet Ortho pix4d Filtering DTM pix4d

12 Workflow Reference data Images Helimap Trajectory GPS-IMU GCP BA Bingo-F SocetSet APM Tie/GCP points LiDAR Helimap Image orientation Camera parameters Point cloud SocetSet mosaic DTM ref Filtering Ortho ref

13 Camera calibration Bingo Camera calibration with Bingo Important tangential distortion (~20 m) Impossible to remove completely non radial deformation

14 Camera calibration Pix4D Camera calibration Pix4D SocetSet/BINGO Pix4D Focal c [mm] PPA X/Y [mm] / / Not the same sensor model

15 AT results 0 [µm] RMS of object points [mm] RMS GCP residuals [mm] X Y Z X Y Z AT-Socet-refdata AT- Socet-swinglet AT- Pix4D-swinglet * na na na 20 (combined) * Corresponds to average magnitude of residuals, not sigma of image residuals Not enough statistics on Pix4D Bundle adjustment for comparison Accuracy of ½ pixel for swinglet, 1/3 of pixel for Ref. data

16 AT results Socet/Bingo Ground Tie points compared to LiDAR DTM Discrepencies with distance from GCP s

17 DTM comparaison Pix4D-LiDAR filtered DTM difference on hard surface < 20cm High differences corresponds to vegetation area or building

18 DSM comparaison Pix4D-LiDAR filtered DSM difference on hard surface < 20cm High differences corresponds to terrain discontinuities and vegetatio area

19 DTM comparaison NGATE-LiDAR Filtered DTM Noise on DTM High difference while distance increases from GCP s area

20 DTM summary (GCP s) [m] Average Standard. deviation Min Max LiDAR Pix4D NGATE Height differences statistics between GCP s and DTM s

21 Orthophoto Reference Swinglet-Pix4D Swinglet Socet-NGATE

22 Results: Ortho Feature line/gcp Red: reference Green: Pix4D Pink: Socet-NGATE GCP residuals [m] Socet set NGATE Pix4D X Y X Y Average RMS

23 Conclusion The Swinglet is an interesting mapping toy flexible no real air regulations for this range of weight (until when? What else in other countries?) designed for small areas Low cost mid range accuracy mapping (20cm)

24 Conclusion Mapping performance depends on processing scheme Conventional softwares reach limitations with such poor quality camera accuracy of 150m AGL. Inaccurate Stereoplotting (jumps between models) New generation algorithm provides better results on DTM/DSM accuracy of ~15cm Ortho in both case is affected by artefacts (DTM, mosaic, jpeg). Planimetric accuracy of ~10-15cm

25 Outlooks New improvement of Pix4D to be tested (dense matching DSM, GCP input ) Other algorithm : Open Source MicMac from IGN Deep noise analysis of DTM? When the professional software will wake up?

26