The Internet of Things: Roadmap to a Connected World. IoT and Localization

Transcription

1 IoT and Localization Daniela Rus Andrew (1956) and Erna Viterbi Prof., EECS Director, CSAIL Computer Science and Artificial Intelligence Laboratory (CSAIL) Massachusetts Institute of Technology

2

3 4x 4x 4x 2x

4

5

6 Rico Shen 2008

7 OUTLINE Introduction to Localization Robust Localization Outdoors Localization Indoors Localization

8 LOCALIZATION PROBLEM STATEMENT Given some representation of the environment, to localize, device must, through sensing, determine its pose with respect to the specified representation Pose = (location, heading) wrt external frame Global coordinate frame E.g., GPS (Earth) coordinates Local coordinate frame Ceiling or floor tiles Mission starting pose Environment features E.g., nearby walls, corners, markings y x

9 OPEN LOOP POSE ESTIMATION Estimate pose from expected results of motion No sensing Dead reckoning: Use odometry to estimate pose w.r.t. initial coordinate frame Multiple error sources; Pose error accumulates with time and motion

10 OPEN LOOP POSE ESTIMATION

11 LOCALIZATION SCENARIOS Estimating location in 2D From measured ranges (distances) From measured bearings (directions) L L P L L L q L L

12 SENSORS AND MEASUREMENTS Range to surface patch, corner Sonar, IR Bearing (absolute, relative, differential) Compass; vision (calibrated camera) Range to point RSS, TOF from RF/acoustic beacon TDoA of acoustic & RF pulse Range and (body-relative) bearing to object Radar return Laser range scanner return Vision (stereo camera rig)

13 TRIANGULATION Natural geometry for 2D localization Simplest framework combining range, bearing Used by Egyptians, Romans for engineering L L L P Range d (distance from from P to L) L q L L L Bearing q (relative to straight ahead in robot frame)

14 TRIANGULATION FROM RANGE DATA Device at unknown position P measures distances d 1, d 2 to known landmarks L 1, L 2 Given d 1, d 2, what are possible values of P? L 1 d 1 y? P L 2 d 2 Robot measures d 1, d 2 x

15 TRIANGULATION FROM RANGE DATA Device must lie on circles of radius d 1, d 2 centered at L 1, L 2 respectively L 1 d 1 L 2 y d 2 P x

16 TRIANGULATION FROM RANGE DATA Change basis: put L 1 at origin, L 2 at (a,0) y L 1 = (0,0) x = (a 2 + d d 2 2 ) / 2a y = ± (d x 2 ) d 1 L 2 = (a, 0) P (Try e.g. setting d 1 = a, d 2 = 0) d 2 x

17 TRIANGULATION FROM RANGE DATA Change basis: put L 1 at origin, L 2 at (a,0) y L 1 = (0,0) x = (a 2 + d d 2 2 ) / 2a y = ± (d x 2 ) d 1 L 2 = (a, 0) P (Try e.g. setting d 1 = a, d 2 = 0) d 2 x Are we done?

18 TRIANGULATION FROM RANGE DATA Two solutions in general, P and P How to select the correct solution? L 1? P? d 1 P d 2 L 2

19 DISAMBIGUATING SOLUTIONS A priori information (richer map) L 1 P d 1 L 2 P d 2

20 DISAMBIGUATING SOLUTIONS Continuity (i.e., spatiotemporal information) L 1 P d 1 L 2 P d 2 Position now Position 10 minutes ago

21 DISAMBIGUATING SOLUTIONS Additional landmarks (redundancy) L 1 P L 3 d 1 d 3 L 2 P d 2

22 TRIANGULATION FROM RANGE DATA Are we done yet, i.e., is pose fully determined? No: absolute heading is not determined L 1 P d 1 L 2 P d2?

23 TRIANGULATION FROM BEARING DATA Body-relative bearings to two landmarks Bearings measured relative to straight ahead Robot observes: L 1 at bearing q 1 L 2 at bearing q 2 L 2 q 2 a L 1 differential bearing a = q 2 q 1 q 1 q = 0 (radians)

24 TRIANGULATION FROM BEARING DATA Body-relative bearings to two landmarks Bearings measured relative to straight ahead Robot observes: L 1 at bearing q 1 L 2 at bearing q 2 L 2 q 2 a L 1 differential bearing a = q 2 q 1 q 1 q = 0 (radians) are two bearings enough for unique localization?

25 Triangulation from two bearings q 1 q 2 L 2 L 1 a a a a Device somewhere on black circular arc Can it be anywhere on circle? (No; ordering constraint)

26 TRIANGULATION FROM BEARING DATA Measure bearing to third landmark Yields robot position and orientation Also called robot pose (in this case, 3 DoFs) L 1 L 2 L 3

27 MEASUREMENT UNCERTAINTY: RAGES Ranges, bearings are typically imprecise Range case (estimated ranges ~d 1, ~d 2 ) L 1 ~d 1 ~d 2 L 2 P Locus of likely positions

28 MEASUREMENT UNCERTAINTY: BEARING L 1 L 2 Two-bearing case (estimated bearings ~q 1, ~q 2 ) What is locus of recovered vehicle poses? Solve in closed form? Is there an alternative?

29 MEASUREMENT UNCERTAINTY Bearing case (measurements ~q 1, ~q 2, ~q 3 ) L 1 L 2 L 3 {P, q} is this always a satisfactory Locus pose of bound? likely poses

30 LOCALIZATION SUMMARY Range based localization Bearing base localization Uncertainty

31 OUTLINE Introduction to Localization Robust Localization Outdoors Localization Indoors Localization

32 LOCALIZATION WITH NOISY RANGES (-1,12) Characteristics: Robust against noise No beacons Handles mobility (-4,7) (-7,1) cm (10,0) (-5,-5) (6,-5) D. Moore, J. Leonard, D. Rus, S. Teller, The Robust Internet distributed of Things: Roadmap network to a localization Connected World with noisy range measurements, Proceedings of the 2004 ACM International Conference on Embedded Networked Sensor Systems (Sensys 2004)

33 COMPLICATIONS OF NOISE Small measurement errors due to noise lead to large localization errors Example: flip ambiguity from noise A A D C OR C D B Small error in CD leads to large position error of D B D. Moore, J. Leonard, D. Rus, S. Teller, Robust distributed network localization with noisy range measurements, Proceedings of the 2004 ACM International Conference on Embedded Networked Sensor Systems (Sensys 2004)

34 THE ROBUST QUADRILATERAL Consider this graph: Robustness characteristics: Rigid (no continuous deformations) No discontinuous flex ambiguities (by Laman s Theorem) We probabilistically constrain it to minimize the likelihood of a flip ambiguity We call it a robust quadrilateral A graph constructed from overlapping robust quads will itself possess the robustness characteristics

35 TRILATERATION W/ ROBUST QUADS If three nodes of a quad have known position, fourth can be computed with trilateration Quads can be chained in this manner

36 COMPLETE ROBUSTNESS TEST Test for robust triangle: Let b = minimum edge length of the triangle Let θ = minimum angle of the triangle Then, in worst case: d err = b sin 2 θ If b sin 2 θ > 3σ then probability of a flip is less than 1% Test for robust quadrilateral: Quad is robust if all four sub-triangles are robust

37 ALGORITHM Starting cluster w/ distance measurements Choose two neighboring nodes for initial robust triangle (-5,-5) (0,0) (6,-5) (0,0) (0,12) x (10,0) y

38 ALGORITHM (CONT.) Cluster localization complete (-5,-5) (6,-5) (0,0) (-4,7) (10,0) (0,12)

39 Distance (cm) Localization error (cm) EXPERIMENTAL RESULTS Robot path over time Error in robot path over time Distance (cm) Time (seconds) (Ground truth derived from computer vision algorithm applied to captured video)

40 Distance (cm) SIMULATION RESULTS With robust quads Localized position versus ground truth Without robust quads Noise std. dev. σ = 5.0 cm Communication radius MSE = 16.2 cm MSE = 54.9 cm 75% of nodes localized 99% of nodes localized Distance (cm) Distance (cm)

41 Localizing the left-over nodes

42 Robot broadcasts locations

43 Robot broadcasts locations

44 OUTLINE Introduction to Localization Robust Localization Outdoors Localization Indoors Localization

45 No Landmarks But Map

46 GPS

47 Tilted-down LIDAR Planar LIDAR Webcam IMU+Odom 2016 Massachusetts Institute 53/56 of Technology

48 2016 Massachusetts Institute 54/56 of Technology

49 Curb Vertical features Surface 2016 Massachusetts Institute 55/56 of Technology

50 2016 Massachusetts Institute 56/56 of Technology

51 Curb features 2016 Massachusetts Institute 57/56 of Technology

52 Vertical Surface 2016 Massachusetts Institute 58/56 of Technology

53

54 MONTE CARLO LOCALIZATION FOR POSE 60/56

55 CURB-INTERSECTION FEATURE EXTRACTION 2D LIDAR 61/56

56 CURB-INTERSECTION FEATURE EXTRACTION Filter 62/56

57 CURB-INTERSECTION FEATURE EXTRACTION Filter 63/56

58 VIRTUAL SENSORS FOR CURB AND INTERSECTION Virtual LIDAR of Curb features (accumulates and fuses points) Likelihood model Virtual LIDAR of intersection features (triggered when there are no curb points) Beam model 64/56

59 EXPERIMENTAL EVALUATION OF MCL Localization Results Curb Map Odometry GPS/INS Detected Curbs MCL Method Curb Outline 66/56

60 LOCALIZATION ACCURACY Odometry GPS/INS MCL Method Curb The Internet Outlineof Things: Roadmap to a Connected World 67/56

61 ZJ Chong, B. Qin, T, Bandyopadhyay, M. Ang, E. Frazzoli, D. Rus, Synthetic 2D LIDAR for Precise Vehicle Localization in 3D Urban Environment, Proceedings of the 2013 Int. Conf. on Robotics and Automation (ICRA), 2013 IEEE International Conference, pp Massachusetts Institute 68/56 of Technology

62 Curb-intersection features exist Uncertainty on straight segments 69/56

63 Curb-intersection features exist Uncertainty on straight segments Use Vertical Surfaces 70/56

64 2016 Massachusetts Institute 71/56 of Technology

65 3D ROLLING WINDOW FOR FEATURES 72/56

66 2016 Massachusetts Institute 73/56 of Technology

67 Surface normal For classification 2016 Massachusetts Institute 74/56 of Technology

68 2016 Massachusetts Institute 75/56 of Technology

69 2016 Massachusetts Institute 76/56 of Technology

70 2016 Massachusetts Institute 77/56 of Technology

71 Synthetic LIDAR based localization The Internet of Things: Roadmap to Lateral a Connected World 2016 Massachusetts Institute 78/56 of Technology

72 ZJ Chong, B. Qin, T, Bandyopadhyay, M. Ang, E. Frazzoli, D. Rus, Synthetic 2D LIDAR for Precise Vehicle Localization in 3D Urban Environment, Proceedings of The the Internet 2013 of Int. Things: Conf. Roadmap on Robotics to a Connected and Automation World (ICRA), 2013 IEEE International Conference, pp Massachusetts Institute 79/56 of Technology

73 OUTLINE Introduction to Localization Robust Localization Outdoors Localization Indoors Localization

74 INDOOR GPS TODAY? Businesses Research [RADAR 00], [Cricket 04], [Lease 04], [Horus 05], [PlaceLab 05], [Beepbeep 07], [SrndSense 09], [EZ 10], [Compac 10],[BatPhone 11], [Wigem 11], [Zee 12], [Will 12],[Centaur 12], [Unloc 12], [PinLoc 12], [FootSlm 12], [ArrayTrack 13], [Guoguo 13], [PinPoint 13],...

75 S. Kumar, S. Gil, D. Katabi, D. Rus, Accurate Indoor Localization with Zero Startup Cost, Proceedings of Mobicom 2014.

76 S. Kumar, S. Gil, D. Katabi, D. Rus, Accurate Indoor Localization with Zero Startup Cost, Proceedings of Mobicom 2014.

77 Pos 1 Pos 2 Localization Software

78 WITH MANY ANTENNAS Such a device does not exist!

79 SYNTHETIC APERTURE RADAR (SAR) Move device to emulate antenna array

80 Users to rotate tablet Process signals as circular array

81 EVALUATION Large Library 5 Access Points Baseline: Angle-of-Arrival The Internet Things: using Roadmap 2-Antenna to a Connected World Array

82 LOCALIZING OBJECTS OF INTEREST AROUND YOU To: Bob Found interesting (21.39, 35.13)

83 OBJECT GEO-TAGGING To: Facilities Reporting Broken (30.56, 21.23)

84 EVALUATION OF OBJECT GEOTAGGING Integrate Ubicarse with VSFM toolkit Localize books in the library

85 SUMMARY Localization as pervasive, seamless, instantaneous service Localized machines will be smarter and more engaged with us Localized devices will help us know more about us & world

86 CDF ACCURACY IN GEO-TAGGING 17 cm Error in Dimension (m)

87 THANK YOU! Daniela Rus Andrew (1956) and Erna Viterbi Prof., EECS Director, CSAIL Computer Science and Artificial Intelligence Laboratory (CSAIL) Massachusetts Institute of Technology