Andreas Birk - 3D mapping in marine environments
Transcription
Andreas Birk - 3D mapping in marine environments
3D Mapping in Marine Environments Andreas Birk Jacobs University http://robotics.jacobs-university.de Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 1 Jacobs University Robotics Group Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 2 Jacobs Robotics especially • 3D Perception (e.g., obstacle avoidance, object recognition) • 3D Worldmodels (e.g., maps – plus adding semantics) in unstructured environments • e.g., marine applications, safety/security/rescue, logistics, … Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 3 3D Mapping Jacobs University, Robotics Group 3D Perception & Semantic Maps http://robotics.jacobs-university.de/ 4 Jacobs Robotics • increasingly active in marine robotics - due to development of methods that can cope with - noisy, unreliable sensor data (noise + outliers) - in unstructured scenes (clutter, occlusions, limited a priori knowledge, etc.) • running marine robotics projects - EU FP7 MORPH EU FP7 CADDY EU H2020 DexROV two national projects Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 5 EU FP7:Marine robotics system of self-organizing logically linked physical nodes (MORPH) Key MORPH concept: • self-reconfiguring robot • consisting of several closely coupled vehicles • for operations in complex 3D marine environments Partners • ATLAS Elektronik • Consiglio Nazionale delle Ricerche • IFREMER • Jacobs University • Instituto Superior Tecnico • Ilmenau University of Technology • CMRE • Universitat de Girona • Institute of Marine Research, IMAR Jacobs • online 3D mapping • view-planning Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 6 EU FP7: Cognitive autonomous diving buddy (CADDY) Robot as Divebuddy & Assistent • Human-Machine-Interaction under challenging conditions • applications (endusers): Archeology, Search & Recovery Partners • University of Zagreb • Consiglio Nazionale delle Ricerche • Instituto Superior Tecnico • Jacobs University • University of Vienna • Newcastle University • Diver Alert Network Europe Jacobs • hand/diver detection & segmentation • mapping as service task Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 7 EU H2020: Effective Dexterous ROV Operations in Presence of Communications Latencies (DexROV) Intelligent Supportfunctions for Teleoperation e.g., in Oil- & Gasproduction (1.5 – 2.5km depth) Jacobs Robotics • recognize & track objects in 3D • semantic 3D mapping Jacobs University, Robotics Group http://robotics.jacobs-university.de/ Partners • Space Applications Services • Comex • UNIGE-ISME • Jacobs University • Idiap Research Institute • Graal Tech • EJR-Quartz 8 3D Mapping in Marine Environments Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 9 Background: Simultaneous Localization and Mapping (SLAM) chicken & egg problem • build map, which needs localization • while using map for localization main idea • roughly localize by - vehicle motion estimation with odometry / navigation sensors - plus local sensor data association, aka registration • optimize localizations/map after registrations to previously visited places (loop closing) to minimize the cumulative error Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 10 Overview on the rest of this talk 4 Main Parts • intro: proper meaning of 3D • marine "3D" sensors • 6-DOF registration (aka SLAM front-end) • short notes on map optimization (aka SLAM back-end) Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 11 Challenges for 3D Mapping in Marine Applications Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 12 3D Map??? No!!! Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 13 3D: Very Challenging & Important Bathymetrie is not 3D!!! • but 2.5D • i.e., 2D manifold in 3D space • can not represent 3D objects / scenarios e.g.: • cliffs (biomonitoring) • oil- & gasindustry • offshore windparks • shipwrecks • harbors © NOCS February 2007 e.g., MORPH : true 3D operations by AUVs getting 3D data is hardly possible from the surface: you need to explore in 3D!!! Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 14 Challenges in 3D Underwater Mapping • real 6 degrees of freedom (DOF) • localization limited (and costly) • limitations in range sensors this all makes registration challenging 6 DOF not really 6 DOF 3D data but motion on a plane (3 DoF) EU-project MORPH Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 15 Marine "3D" Sensors Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 16 “3D” Sensors • all "3D" sensors deliver 2.5D range - can be represented as range image, i.e., array of ranges - this can be exploited for processing - real 3D only through registration of multiple scans!!! • different native data formats for scans - often even directly in 3D, e.g., point cloud - projecting back to 2.5D then requires sensor model Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 17 Marine 3D Sensors • actuated 2D Multibeam Echo Sounder Blueview BV5000 • solid state 3D Multibeam Tritech Eclipse • structured light • stereo vision • actuated Laser Range Finder Jacobs University, Robotics Group 3D-at-Depth SL-1 http://robotics.jacobs-university.de/ 18 Marine 3D Sensors • actuated 2D Multibeam Echo Sounder Blueview BV5000 • solid state 3D Multibeam Tritech Eclipse • structured light • stereo vision • actuated Laser Range Finder Jacobs University, Robotics Group 3D-at-Depth SL-1 http://robotics.jacobs-university.de/ 19 Underwater Stereo Vision • underwater & stereo: old technique - Jean de Wouters d'Oplinter, 1948 stereo tests at Cote D’Azur - Dimitri Rebikoff, 1954 stereo for mapping archeological sites payload on manned vehicle Pegasus • also small own development - standard algorithms for dense stereo (including CUDA version on GPU) in real time in self-contained system research contribution: fast registration Max Pfingsthorn, Heiko Bülow, Igor Sokolovski, Andreas Birk. Underwater Stereo Data Acquisition and 3D Registration with a Spectral Method. IEEE Oceans, Bergen, Norway, 2013 Jacobs University, Robotics Group input: 2D images http://robotics.jacobs-university.de/ output: 3D colored point cloud 20 Calibration & Rectification with Pinax Model • good calibration essential for good stereo data • in general, current state of the art for calibration is flawed • flat pane interface - not a pinhole camera anymore - but axial camera • water refraction index - usually ignored - but significant influence of salinity - if data recorded at different place than calibration, severe errors possible Jacobs University, Robotics Group from: Treibitz, T.; Schechner, Y.Y.; Kunz, C.; Singh, H., Flat Refractive Geometry. Pattern Analysis and Machine Intelligence, IEEE Transactions on , vol.34, no.1, pp.51,65, Jan. 2012 http://robotics.jacobs-university.de/ 21 Calibration & Rectification with Pinax Model • under realistic design constraints (camera front close to flat pane) • combination of Pinhole with Axial model (PinAx) - using Axial 12th degree polynomial projection function - to generate virtual pinhole correction via look-up table (very fast!!!) - water refraction index (salinity) as parameter (estimated or from CTD) • calibration simply once in-air Tomasz Luczynski, Max Pfingsthorn, Andreas Birk. The Pinax-Model for Accurate and Efficient Refraction Correction of Underwater Cameras in Flat-Pane Housings. (under review) Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 22 3D Sonar: Tritech Eclipse sensing parameters: • • • • • • • • • Operating Frequency: 240 kHz Beam Width: 120 deg Number of Beams: 256 Acoustic Angular Resolution: 1.5 deg Effective Angular Resolution: 0.5 deg Depth/Range Resolution: 2.5 cm Maximum Range: 120 m Minimum Focus Distance: 0.4 m Scan Rate: 140 Hz at 5 m, 7 Hz at 100 m physical parameters: • • • • • • • Width: 342 mm Height: 361 mm Depth: 115 mm Weight Wet / Dry: 9 kg /19 kg Depth Rating: 2500 m Power Consumption: 60 W Supply Voltage Nominal: 20-28 VDC Jacobs University, Robotics Group pictures courtesy of Tritech, UK http://robotics.jacobs-university.de/ 23 Lesumer Sperrwerk 17 scans, ca. 110m x 70m scan 1 (start angle) scan 4 scan 17 (end angle) Google earth Jacobs University, Robotics Group scan 17 http://robotics.jacobs-university.de/ 24 3D without Range Sensors there are techniques for full 3D from just 2D camera data (video) • Structure from Motion (SfM) / Multi-View Registration / Bundle Adjustment • sometimes aka photogrammetry disadvantages • computational complexity (offline) • non-metric Jacobs University, Robotics Group (just one) example: T. Nicosevici, N. Gracias, S. Negahdaripour, and R. Garcia, "Efficient three-dimensional scene modeling and mosaicing," Journal of Field Robotics, vol. 26, pp. 759-788, 2009. http://robotics.jacobs-university.de/ 25 Registration Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 26 Registration problem: • given two sensor data sets • find parameters of transform to spatially align them, • i.e., 3-DoF (2D) or 6-DoF (3D) • (including uncertainty estimates for SLAM) Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 27 27 Standard for 3D Registration: Iterative Closest Point (ICP) Assume correspondences by nearest neighbors • kd-tree for efficiency Given correspondences • Horn's algorithm • for closed form least squares fit • using quaternions to get the rotation part iterate until convergence Jacobs University, Robotics Group Besl, P. J., & McKay, N. D. (1992). A method for registration of 3-D shapes. IEEE Trans. on Pattern Analysis and Machine Intelligence (Vol. 14, pp. 239-256). Zhang, Z. (1994). Iterative point matching for registration of free-form curves and surfaces. Int. J. Comput. Vision, 13(2), 119-152. doi: 10.1007/bf01427149. Horn, B. K. P. (1987). Closed-form solution of absolute orientation using unit quaternions. Journal of the Optical Society of America, 4(4), 629-642. many variants possible, e.g. point to plane Rusinkiewicz, S., & Levoy, M. (2001). Efficient variants of the ICP algorithm. Paper presented at the 3-D Digital Imaging and Modeling. Proceedings. Third International Conference on. nice "full package" (i.e., incl. SLAM back end and visualization) Nuechter, A., 3D Robotic Mapping. Springer Tracts in Advanced Robotics (STAR). Springer. 2009 http://robotics.jacobs-university.de/ 28 Limits of ICP & co • needs good initial starting conditions - excellent navigation needed - hard, respectively costly in underwater applications • is challenged by partial overlap and noise example of failed ICP registrations • Lesumer Sperrwerk dataset • with Tritech Eclipse data Heiko Bülow and Andreas Birk. Spectral Registration of Noisy Sonar Data for Underwater 3D Mapping. Autonomous Robots, 30 (3), pp. 307-331,Springer, 2011 Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 29 There is more than ICP & co… • nice and good starting points - especially, easily understandable for CS-people ☺ - very efficient, highly tuned implementations exist • but - they rely on very local information hence very sensitive to only partial overlap and disturbances, especially dynamics - and are iterative methods that require good starting conditions => use overall (global) appearance of scans instead Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 30 "Global" Appearance: e.g. Dominant Planes 3D Plane SLAM 1. consecutive acquisition of 3D range scans 2. extraction of planes including uncertainties 3. registration of scans based on plane sets • determine correspondence set • • find the optimal decoupled • • 4. 5. maximizing the global rigid body motion constraint rotations (Wahba's problem) and translations (closed form least squares) embedding in a pose graph loop detection & relaxation (SLAM proper) Jacobs University, Robotics Group Kaustubh Pathak, Narunas Vaskevicius and Andreas Birk. Uncertainty Analysis for Optimum Plane Extraction from Noisy 3D Range-Sensor Point-Clouds. Intelligent Service Robotics, Vol.3, Iss.1, p.37-48, Springer, 2010 Kaustubh Pathak, Andreas Birk, Narunas Vaskevicius, and Jann Poppinga, Fast Registration Based on Noisy Planes with Unknown Correspondences for 3D Mapping, IEEE Transactions on Robotics, 26 (3), pp. 424 – 441, 2010 K. Pathak, A. Birk, N. Vaskevicius, M. Pfingsthorn, S. Schwertfeger, and J. Poppinga. Online 3D SLAM by Registration of Large Planar Surface Segments and Closed Form Pose-Graph Relaxation. Journal of Field Robotics, Spec.Iss. on 3D Mapping, 2010 http://robotics.jacobs-university.de/ 31 Planes reasonable in Unstructured Environments European Space Agency (ESA) Lunar Robotics Challenge Teide Volcanic Crater Tenerife, Spain, 2008 Narunas Vaskevicius, Andreas Birk, Kaustubh Pathak, and Soeren Schwertfeger. Efficient Representation in 3D Environment Modeling for Planetary Robotic Exploration. Advanced Robotics, Vol. 24, Iss. 8-9, pp. 1169-1197, Brill, 2010 Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 32 Plane based 6 DoF registration by MUMC infinite plane: normal dist. to origin plus uncertainty: covaricance • “left” and “right” view • two plane sets • find best correspondences - to minimize uncertainty under rigid body motion constraints • without combinatorial explosion - start with pairs & optimize rot./trans. max. consensus set with min. uncert. Jacobs University, Robotics Group Minimum Uncertainty Maximum Consensus (MUMC) Kaustubh Pathak, Andreas Birk, Narunas Vaskevicius, and Jann Poppinga, Fast Registration Based on Noisy Planes with Unknown Correspondences for 3D Mapping. IEEE Transactions on Robotics, 26 (3), pp. 424 – 441, 2010 http://robotics.jacobs-university.de/ 33 Plane-Registration on Sonar Data Lesumer Sperrwerk Dataset • good correspondence with ground truth • fringe benefit: high compression K. Pathak, A. Birk, and N. Vaskevicius. Plane-Based Registration of Sonar Data for Underwater 3D Mapping. International Conference on Intelligent Robots and Systems (IROS), Taipeh, Taiwan, 2010, pp. 4880 - 4885. Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 34 Spectral Registration with Multilayer Resampling (SRMR) • Fast Fourier Transform (FFT) on discretized 3D range data • decouple translation and rotation and • resample for spectral registrations in 1D, 2D, 3D to get all 6 DOF main aspects: • use of use of phase only matched filters (POMF) • process whole stack of spherical layers in one step advantages • fast fixed computation time • very robust against noise • works with partial overlap and significant occlusions Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 35 Spectral Registration with Multilayer Resampling (SRMR) sketch of SRMR algorithm • yaw determination: - • roll-pitch determination: - • resample hemispheres (projection in spherical coordinates) on different radii from the magnitude of the 3D spectrums determine the yaw angle by a rotational registration (polar resampled) from the resampled structures (3D POMF) re-rotate the 3D spectrum according to the determined yaw angle in order to align the spectrums for yaw resample hemispheres (rectangular projection) on different radii from the magnitude of the 3D spectrums determine roll and pitch angle by translational registration from the resampled structures (3D POMF) translational registration: - re-rotate scan data according to all determined angles in order to align the scans for the remaining 3D translation determine the 3D translation between the rotationally aligned scans by a 3D POMF registration H. Bülow and A. Birk. Spectral 6-DOF Registration of Noisy 3D Range Data with Partial Overlap. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), vol. 35, pp. 954-969, 2013. Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 36 Spectral Registration with Multilayer Resampling (SRMR) example results • underwater sonar: Lesum river lock • Stanford bunny • Bremen downtown dataset • SSRR collapsed car park Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 37 Collapsed Car-Park dataset collected at • 2008 Response Robot Evaluation Exercise • in Disaster City, College Station, Texas main sensor: actuated Laser Range Finder • SICK S 300 • plus cheap toy servo Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 38 Collapsed Car-Park: Very Hard for 3D Mapping • large motions of the robot, i.e., small overlap between scans • with real 6 DoF, i.e., simultaneous change of roll, pitch, yaw • without usable motion estimates - odometry completely off - gyro severely affected by vibrations Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 39 Performance of ICP (not only tested by ourselves but also by experts ☺ ) red dashed lines indicated multiple locations of the facade in ICP-„registered“ scan pairs ICP on Crashed Car Parking set: failure for 14 out of 25 scan pairs Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 40 Comparison of Reg-Methods on Crashed Car-Park run times comparison results: 1. SRMR 2. SOFT spectral 3. plane registration 4. ICP 5. principle axis 6. HSM3D success rates Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 41 SRMR on Sonar Data • again: very robust • and fast (with fixed computation time) scan 1 (start angle) scan 17 (end angle) Heiko Bülow and Andreas Birk. Spectral Registration of Noisy Sonar Data for Underwater 3D Mapping. Autonomous Robots, 30 (3), pp. 307-331,Springer, 2011 Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 42 Simultaneous Localization and Mapping (SLAM) Back-end: Loop Closing and Optimization Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 43 Simultaneous Localization and Mapping (SLAM) main idea • roughly localize by navigation and registration • optimize localizations/map by registration with previously visited places (loop closing) and minimize the cumulative error now (very short) glimpse on 2nd aspect Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 44 Loop Closing • proximity based - easiest possible strategy: use current localization estimate - to check whether there are previously visited places around - if so: try registrations • place recognition - non-trivial, especially with respect to good strategies to get reasonable cost/benefit - typically (visual or 3D) feature collections in associative representation (hashes) Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 45 Place Recognition • e.g., FAB-MAP - monocular images - based on bag of words on SURF - very efficient (linear in #places) Cummins, M., & Newman, P. FAB-MAP: Probabilistic Localization and Mapping in the Space of Appearance. The International Journal of Robotics Research, 27(6), pp. 647-665. 2008 Cummins, M., & Newman, P. Appearance-only SLAM at large scale with FAB-MAP 2.0. The International Journal of Robotics Research, 30(9), pp. 11001123. 2011 • extension to stereo - both 2D visual features (SURF) - plus shape features Ivaylo Enchev, Max Pfingsthorn, Tomasz Luczynski, Igor Sokolovski, Andreas Birk, Daniel Tietjen. Underwater Place Recognition in Noisy Stereo Data using Fab-Map with a Multimodal Vocabulary from 2D Texture and 3D Surface Descriptors. IEEE Oceans. Genoa, Italy, 2015 Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 46 SLAM: optimization part use loop closures as "feedback" to minimize error • "historically": Kalman Filter SLAM - n features => n2 variables - zero-mean white Gaussian noise assumed - data association (how to match features) • hence alternatives popular - e.g., particle filter aka “condensation”, “(sequential) Monte Carlo” or “survival of the fittest“ algorithm - e.g., graph based Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 47 Uncertainty/Confidence Estimates in Registrations for SLAM very useful as "weights" for each registration result to measure the "quality" of the spatial estimates • well established for ICP & co • inherent in plane-registration • but not studied for 3D spectral methods => extension for SRMR registration Max Pfingsthorn, Andreas Birk, and Heiko Buelow, Uncertainty Estimation for a 6-DoF Spectral Registration method as basis for Sonar-based Underwater 3D SLAM, International Conference on Robotics and Automation (ICRA), Saint Paul, Minnesota, IEEE Press, 2012 Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 48 Generalized Graph SLAM handling Ambiguities • Local Ambiguity: - Ambiguity in sequential observations - Spatially and temporally close • Global Ambiguity: - Ambiguity in non-sequential observations, i.e. loops - Spatially close, temporally far Max Pfingsthorn and Andreas Birk. Simultaneous Localization and Mapping (SLAM) with Multimodal Probability Distributions. International Journal of Robotics Research, 32(2), pp. 143-171, Sage, 2013 Max Pfingsthorn and Andreas Birk. Generalized Graph SLAM: Solving Local and Global Ambiguities through Multimodal and Hyperedge Constraints. International Journal of Robotics Research (IJRR). Sage, 2015 Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 49 Local Ambiguity or can arise from complementary motion estimates: one fails => mutually exclusive choices better option than probabilistic fusion Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 50 Global Ambiguity example: Loop Detection place recognition, here with FabMAP, can give multiple results => again: need for representing alternative motion options Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 51 Representation as Multimodal Hypergraphs mid-stage: SLAM front end multimodal hypergraph Generalized Prefilter (discrete optimization) standard pose-graph SLAM back end • new discrete optimization stage between front- and back-end • generate standard graph without ambiguity find most globally consistent component combination by traversing a spanning tree with heuristics to fight curse of combinatorial explosion Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 52 Generalized Graph SLAM multiple alternative spatial relations in one probability density: • local ambiguity as Mixture of Gaussians (MoG) • global ambiguity as mixture of constraints in a hyperedge from registration from loop detection hyperedge weights Jacobs University, Robotics Group one MoG per hypothesis in hyperedge http://robotics.jacobs-university.de/ 53 Results of Generalized Graph SLAM vs state of the art methods Sünderhauf’s Switchable Constraints Agarwal’s Dynamic Covariance Scaling (DCS) Olson’s MaxMixture Latif’s Realizing, Reversing, Recovering (RRR) Generalized Graph SLAM with Prefilter Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 54 Short Final Note also work on how to • plan to efficiently generate the maps (exploration) • ensure good coverage (view-planning) • make sense of them (object/terrain classification & semantic mapping) Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 55 Conclusions • bathymetry is not 3D!!! • 6 DoF registration - complementing (or even replacing) navigation there is more than ICP there are alternatives that are much more robust (and faster) e.g., plane based registration with MUMC e.g., spectral registration with SRMR • SLAM back-end - Generalized Graph SLAM - can handle outliers in local motion estimates and loop closures - e.g., short loss of navigation or ambiguous place recognition • efficient map generation & making sense of them Jacobs University, Robotics Group http://robotics.jacobs-university.de/ 56