Integrated Positioning The Challenges New technology More GNSS satellites New applications Seamless indoor-outdoor More GNSS signals personal navigati

Transcription

1 Integrated Indoor Positioning and Navigation Professor Terry Moore Professor of Satellite Navigation Nottingham Geospatial Institute The University of Nottingham

2 Integrated Positioning The Challenges New technology More GNSS satellites New applications Seamless indoor-outdoor More GNSS signals personal navigation Communications 2011 Intelligent NGI Transport Systems WiFi / RFID Rail signalling UWB, Sparse Band Precision aircraft landing Digital broadcasting Ships in harbours Pseudolites, Locatalites Location-dependent billing Smaller, cheaper inertial sensors New mapping (outdoor & indoor) More processing power Drives new applications and virtual security fences Tracking people/animals/assets Social inclusion Creates new challenges

3 Multi-Sensor Positioning Communications Deadreckoning GNSS Different technologies sensors solve different parts of 2011 the positioning NGI problem Terrestrial Featurematching More information radio Positioning navigation sensors enables greater integrity Mapping Diverse technology gives better availability

4 Positioning Integration Optimal position based on all information Increased robustness to interruptions Synergistic 2011 use of sensors NGI GNSS Ground Based Positioning Sensors Signals of Opportunity Dead Reckoning Sensors, INS Seamless Integration Migration between positioning sensors Kalman Filtering Mathematical method of combining data to give optimal position solution, whilst on-the-move

5 Urban eloran Trials

6 Urban eloran Trials

7 Integrated Loran and INS LORAN-C + INS vs LORAN-C Legend Reference LORAN-C + INS LORAN-C only 2011 NGI

8 Digital Audio Broadcasting Signals of Opportunity Designed for dynamic receivers Vehicle entertainment systems Uses high 2011 power terrestrial signals ( NGI 1000 GNSS) Transmitter infrastructure already in place No added cost DAB in the UK: Coverage > 85% (UK population) Digital switchover from FM planned for 2015 Uses Single Frequency Network (SFN) approach Multiple synchronised transmitters per network Receiver switches seamlessly when mobile

9 DAB Positioning Time Difference of Arrival A & B = Synchronised DAB SFN 1 Transmitters C & D = Synchronised DAB SFN 2 Transmitters R = Receiver Location AR BR R A B TDOA1 D TDOA2 C SFN 1 Time SFN 2 Time

10 Initial Positioning Results Test Region 2011 Min Offset NGI 59m G n = GPS Positions D n = DAB Positions = Offset Max Offset 288m

11 Pedestrian Navigation Pedestrian navigation Challenging Walk indoors for long periods of time Prefer 2011 not to install dedicated infrastructure NGI Real-time Inertial navigation Accelerometers (and gyros) on smart phones Cheap = low quality = high position drift Foot mounted sensors promising Integration Need constant updates from other sensors Use GPS when available Other sensors?

12 Integrated Indoor Positioning Infrastructured Use of GNSS with WiFi, RFID, Bluetooth, UWB, etc Dedicated 2011 infrastructure or NGI a priori building information or Ad-hoc networks, drop units or. Signals of opportunity Infrastructureless Use of DR and INS low cost, independent

13 Integrated Indoor Positioning Example Shoe mounted low cost MEMS IMU Zero-velocity update algorithm Using elevator Vertical drift Side View

14 2011 Indoor Navigation NGI using Smartphones Professor Terry Moore Professor of Satellite Navigation Nottingham Geospatial Institute The University of Nottingham

15 Indoor Positioning Background Indoor navigation requirements: Inexpensive Don t want to install dedicated infrastructure High 2011 accuracy (~1m) NGI Ubiquitous Make use of existing technology Smartphones already have technology that can be used for positioning: GPS, Wi-Fi, Mobile network, Bluetooth, Camera, Internet link Wi-Fi already widely used for positioning E.g. Skyhook, Google, Apple

16 Smartphone Sensors GPS Microphone Ambient Wi-Fi light sensor 3G/GPRS Bluetooth Proximity sensor 3-axis accelerometer 3-axis gyro FM radio 3-axis magnetometer Camera

17 Current Positioning Methods GPS Microphone Ambient Wi-Fi light sensor 3G/GPRS Bluetooth Proximity sensor 3-axis accelerometer 3-axis gyro FM radio 3-axis magnetometer Camera

18 Current Positioning Methods Cell-ID Wi-Fi GPS 200m

19 Wi-Fi Fingerprinting GPS Microphone Ambient Wi-Fi light sensor 3G/GPRS Bluetooth Proximity sensor 3-axis accelerometer 3-axis gyro FM radio 3-axis magnetometer Camera

20 Wi-Fi Fingerprinting Makes use of existing Wi-Fi access points Received Signal Strength Indicator (RSSI) RSSI varies spatially Compare 2011 RSSIs from multiple Access Points NGI (APs) to database Don t need to know position of APs Works well inside buildings Typically achieve ~5m accuracy or better with particle filter Survey must record: Received Signal Strength Indicator (RSSI) Position

21 Wi-Fi Fingerprinting How do we record position to build the database? Traditional survey methods time consuming Use map to identify location takes time & prone to error Difficult 2011 to apply over large areas NGI How do we keep the database up-to-date? APs may change (add/replace/remove) Building can change (e.g. remove walls) Furniture can be moved Commercial systems require regular re-survey Proposal Use high accuracy indoor positioning system from NGI Aim to use non-dedicated users to build database e.g. security guards or hospital porters

22 Wi-Fi Fingerprinting Measure signal strengths to all Access Points in view Match measured signal strengths to database Requires database of: Location signal strengths to all Access Points (APs) in view 1.Position, ID1, SS1, ID2, SS2, ID3, SS3,... 2.Position, ID1, SS1, ID2, SS2, ID3, SS3,... 3.Position, ID1, SS1, ID2, SS2, ID3, SS3,... Signal strengths Position 4.Position, ID1, SS1, ID2, SS2, ID3, SS3,

23 WiFi Data Collection Data collected on ground floor of Nottingham Geospatial Building HP laptop 2011 with Wi-Fi NGI Netstumbler for Wi-Fi data collection Foot mounted IMU Microstrain 3DM-GX3 USB comms and power Outputs NMEA data Basic Wi-Fi fingerprinting software developed at NGI

24 WiFi Data Collection Using Foot-mounted IMU Survey Fingerprint trial

25 Signal Strength for one AP Signal strength to AP 00:23:33:16:3C:90 >-40dB -40 to - 50dB -50 to - 60dB -60 to - 70dB

26 Basic Wi-Fi fingerprinting

27 Wi-Fi Fingerprinting Works better indoors where walls/ceilings/furniture will attenuate signals the most Accuracy comes from signal strength varying spatially Advanced 2011 algorithms NGI Particle filtering How do we build databases? Skyhook use fleet of vehicles with GPS (tribe sourcing) Google use crowd sourcing(?) But what about inside where GPS isn t effcetive? Slow database generation using building plans Scalability? How do we keep the database up-to-date? Maintain database using non-dedicated people First responders could build database which others use

28 Inertial Navigation GPS Microphone Ambient Wi-Fi light sensor 3G/GPRS Bluetooth Proximity sensor 3-axis accelerometer 3-axis gyro FM radio 3-axis magnetometer Camera

29 Inertial Navigation 3 gyros and 3 accelerometers Orientation from integrating gyros Displacement from rotating measurements to Earth frame (using gyros), removing gravity and double 2011 integrating accelerometers NGI Not very accurate! MEMS getting better Cheaper (higher volumes e.g. Wii, smartphones) Better manufacturing Calibration Successful results usually from Good sensors Integration with GPS, magnetometers, zero velocity, Step detection algorithms

30 Inertial Navigation Typically suggested for indoor navigation but... Cheap = low quality = large position drift Even navigation grade ~2km/hr Foot mounted sensors promising (not convenient) Accelerometers and gyros now on smart phones Integration Need constant updates from other sensors GPS when available Other sensors? Use Microstrain 3DM-GX3-25 for these examples

31

32 Inertial Navigation Time Horiz (s) error (m)

33 Computer Vision + Inertial Navigation GPS Microphone Ambient Wi-Fi light sensor 3G/GPRS Bluetooth Proximity sensor 3-axis accelerometer 3-axis gyro FM radio 3-axis magnetometer Camera

34 Concept: Video Aided Inertial Phone has: GPS Concept: Video Aided IMU Wi-Fi Camera 3 x Accelerometer 3 x Gyro

35 Concept: Video Aided Inertial Phone contains: GPS Concept: Video Aided IMU Wi-Fi Camera 3 x Accelerometer 3 x Gyro Successive images used to compute translation of the camera Used to correct IMU drift

36 Computer Vision Ground plane Homography algorithm Single camera looking at plane Compute rotation and translation between images 2011 t 1 NGI

37 Computer Vision Algorithm: 1.FAST corner detector About 300 points per image, 9x9 pixels 2.Identify 2011 correspondences using sum-of-squared NGI differences If features are similar all correspondences computed 3. RANSAC/BaySAC algorithm Random subsets of minimum points selected Fit homography model to find inliers 4. Least squares Compute homography from inliers Decompose to compute translation and rotation Feature extraction Identify correspondences BaySAC framework Least Squares

38 Computer Vision Examples... Blue (inlier correspondences) Red (outlier correspondences)

39 INS /Vision Integration Computer vision provides camera frame direction vector and rotation Velocity (scale translation by approximate height / time) Camera 2011 error model: NGI Rotation misalignment (ignore) Sensor axes not co-located (ignore) Scale factor error Observation equations relate body frame velocity to: Camera scale factor error INS navigation frame velocity error INS attitude error

40 INS integration INS mechanisation computes position, velocity and attitude from rotations and accelerations Kalman filter modelling 15 states consisting of: Geodetic 2011 position error NGI Navigation frame velocity error Attitude error Gyro bias Accelerometer bias Use GPS position measurements when available Use translation vector from camera scaled by height Estimate height error in Kalman filter Ignore rotation Ignore axes misalignment between IMU and camera

41 Integration INS corrections Rotation, Acceleration Position, Velocity, Attitude 2011 Kalman IMU INS NGI filter Ranges, Position, Ephemeris Velocity GPS PVT computation Camera Image Computer Vision Translation

42 Experiment GPS Power

43 Experiment Conducted outside so GPS as reference GPS/Inertial u-blox ANTARIS 4 RXMRAW 2011 messages post processed in GrafNav NGI Ordnance Survey reference station < 10km Microstrain 3DM-GX3 0.2 o /s gyro and 0.01g accel bias stability PTDL Camera: Rigidly attached to IMU Handheld Canon Ixus 750 at 30fps Every 4 th frame used 7.5fps Camera approx. looking at ground Power

44

45

46 Position Accuracy Time Horiz (s) error (m)

47 IMU,GPS,Vision Results 7.5Hz Maximum error 4.3m after 6 minutes and 300m travelled

48 INS Processing Requirements Processing time using 3GHz desktop PC POINT integration software (c++) Some inefficiencies due to flexibility of the software Processing Processing Typical time per Function time per update rate second of epoch (ms) (Hz) data (ms) INS mechanisation KF Prediction KF Update Total

49 Computer Vision Processing Requirements Computer vision software developed at UoC (c++) Processing Function time per 2011 epoch (ms) NGI Feature 17 extraction Feature 49 correspondence Homography estimation Homography decomposition 15 Total 86 5

50 Reducing Computer Vision Update Rate Investigate different update rates Camera sample two images close together to get overlap Feature 2011 extraction time therefore doubles NGI Function 7.5Hz 1Hz 0.2Hz INS Computer vision Total Total processing time per 1s of data in ms

51 1.03s update rate

52 5.07s update rate

53 10.00s update rate

54 Maximum Position Error at Different Update Rates Update rate (s) Maximum horizontal error (m)

55 Discussion Demonstrated Computer Vision algorithm: Microstrain 3DM-GX3-25 IMU Significant improvement over IMU only 4.3m max position error instead of 19km after 6 minutes Demonstrated similar performance at 0.13 to 5s update rates Significant reduction in processing requirements Less robust with fewer updates Future work will include: Closer integration between INS and Computer Vision Implementation using smartphone sensors Look at other methods e.g. Particle filtering

56 GPS+IMU+Vision Summary Advantages Good position accuracy Makes use of sensors already on smartphones Handheld Works with or without GPS Disadvantages Needs to be initialised e.g. with GPS Not tested with real smartphones (yet) Problems in low light conditions Computationally expensive

57 Other Research GPS Microphone Ambient Wi-Fi light sensor 3G/GPRS Bluetooth Proximity sensor 3-axis accelerometer 3-axis gyro FM radio 3-axis magnetometer Maps Camera

58 Other Research Magnetometers Total magnetic field varies spatially Image matching (image bag-of-words) Build database of images and locations (like Wi-Fi) Search for an image match to get location Map matching Already used for inertial and Wi-Fi (particle filtering) Walls and doors constrain user motion Direction of travel

59 2011 Indoor Navigation NGI using Foot Mounted IMU Professor Terry Moore Professor of Satellite Navigation Nottingham Geospatial Institute The University of Nottingham

60 Foot Mounted INS NGI have developed a high accuracy indoor positioning system Foot mounted Inertial Measurement Unit (IMU) Zero velocity 2011 update, every step NGI IMU ~ 1700 Requires initialisation on known point Novel heading algorithm used to correct heading errors Shown to consistently maintain <5m accuracy over 40 minutes

61 Unbounded Foot Mounted IMU Shopping Centre

62 Foot Mounted IMU Zero Velocity Measurement (ZUPT) Foot mounted IMU ZUPT every ~0.4 second Lasts around 2011 half of a step NGI Accelleration (ms -1 ) Time (s)

63 ZUPT Corrected Foot Mounted IMU Final Position Error 2011 (~75m) NGI Heading Error in Shop

64 Building Heading aided IMU People in a building tend to move parallel to outside walls of buildings Incorporate this info into the Navigation Filter (EKF) in the form of an observation of heading error.

65 Building Heading & ZUPT Corrected Foot Mounted IMU

66 Foot Mounted IMU ZUPT for 40 mins > 200 m

67 Foot Mounted INS ZUPT and Heading Constraint

68 Multiple Polygons

69 Multiple Polygon Trial 40 mins

70 Conclusions Identified some promising technologies for navigation GPS, Wi-Fi, gyros, accelerometers, magnetometers, cameras, maps Sensors 2011 already available on smartphones NGI Although not necessarily that accurate Non-dedicated infrastructure positioning e.g. Wi-Fi, images, magnetic field What else? Other sensors? Needs a strong case to appear on a smartphone Integration Solution will comprise of several technologies? Use inertial navigation to combine together?

71 Integrated Positioning The Challenges New technology More GNSS satellites New applications Seamless indoor-outdoor More GNSS signals personal navigation Communications 2011 Intelligent NGI Transport Systems WiFi / RFID Rail signalling UWB, Sparse Band Precision aircraft landing Digital broadcasting Ships in harbours Pseudolites, Locatalites Location-dependent billing Smaller, cheaper inertial sensors New mapping (outdoor & indoor) More processing power Drives new applications and virtual security fences Tracking people/animals/assets Social inclusion Creates new challenges

72 Contact Details Professor Terry Moore Director of the NGI Nottingham Geospatial Building The University of Nottingham Triumph Road Nottingham NG7 2TU Telephone: +44 (0) Fax: +44 (0) WWW: