A test bench to improve registration using RGB-D sensors

Transcription

1 A test bench to improve registration using RGB-D sensors François Pomerleau and Stéphane Magnenat and Francis Colas and Ming Liu and Roland Siegwart April 8, European Robotics Forum, Västerås, Sweeden

2 Introduction Current works on icp/registration focus on limited datasets/specific conditions The best solution depends on the application Yet there is a lack of benchmark/comparative studies Solving this requires: Common datasets Versatile implementation 2/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

3 Introduction Current works on icp/registration focus on limited datasets/specific conditions The best solution depends on the application Yet there is a lack of benchmark/comparative studies Solving this requires: Common datasets Versatile implementation 2/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

4 Introduction Current works on icp/registration focus on limited datasets/specific conditions The best solution depends on the application Yet there is a lack of benchmark/comparative studies Solving this requires: Common datasets Versatile implementation 2/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

5 Introduction Current works on icp/registration focus on limited datasets/specific conditions The best solution depends on the application Yet there is a lack of benchmark/comparative studies Solving this requires: Common datasets Versatile implementation 2/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

6 Common datasets Datasets with ground truth, using a RGB-D sensor 3/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

7 Dataset Acquisition Setup Vicon 100 Hz, < 1 mm precision Kinect hand-held, 30 Hz, points Laptop Thinkpad W510, Intel Core i7 Q820 access to FMA/Vicon through the courtesy of Prof. D Andrea 4/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

8 Experiments translations on the three axes, for about 10 s per axis rotations on the three axes, for about 10 s per axis a free fly motion over the scene, for about 15 s slow motion, speed of indoor ground robots (0.3 m/s) medium motion, speed of agile robots (0.5 m/s) fast motion, a challenging speed (1.3 m/s) 5/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

9 Example of Acquisition: Free Fly Video of free fly: slow, medium speed and fast 6/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

10 Versatile implementation Open-source library providing a modular icp chain Tracker 7/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

11 Tracker = Sequence of ICP + Key Frame A 8/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

12 Tracker = Sequence of ICP + Key Frame A 8/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

13 Tracker = Sequence of ICP + Key Frame A B Minimum overlap authorized 8/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

14 Tracker = Sequence of ICP + Key Frame A B Minimum overlap authorized 8/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

15 Tracker = Sequence of ICP + Key Frame A B C 8/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

16 Tracker = Sequence of ICP + Key Frame F A B C D E 8/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

17 Tracking Performances Old desktop PC with an Intel Xeon L5335 Recent laptop with an Intel Core i7 Q 820 Embedded system with an Intel Atom CPU Z530 Minimum number of points Kinect frame rate Frequency (Hz) Number of point used 9/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

18 Modular ICP chain Reading Data filter Data filter Data filter Data filter Data filter Reference Match Iterations Outlier filter Outlier filter Error min. Trans. checker Trans. checker 10/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

19 Typical ICP chain for Office Environments Reading Reference Closest point reduction Random sampling Surface normal extraction Kd-tree Iterations Max distance Median Point-to-plane Nb of iterations Error 11/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

20 Conclusion Ongoing work toward standardisation of icp comparison datasets with ground truth open-source library providing a modular icp chain Tracker as tech demo 12/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

21 Conclusion Ongoing work toward standardisation of icp comparison datasets with ground truth open-source library providing a modular icp chain Tracker as tech demo 12/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

22 Conclusion Ongoing work toward standardisation of icp comparison datasets with ground truth open-source library providing a modular icp chain Tracker as tech demo 12/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

23 Conclusion Ongoing work toward standardisation of icp comparison datasets with ground truth open-source library providing a modular icp chain Tracker as tech demo 12/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors

24 Live Demo and Questions Enjoy! Your questions are welcome! 13/13 Pomerleau et al. A test bench to improve registration using RGB-D sensors