Autonomous Integrity Monitoring of Navigation Maps on board Vehicles - - PowerPoint PPT Presentation

1

• Ph. Bonnifait

Autonomous Integrity Monitoring of Navigation Maps on board Vehicles

Philippe Bonnifait

Professor at the Université de Technologie de Compiègne Heudiasyc UMR 7253 CNRS, France In collaboration with Clément Zinoune and Javier Ibanez-Guzman Renault S.A.S. PPNIV 2015, Hamburg, 28 September 2015

2

• Ph. Bonnifait

Outline Context and Problem Statement Fault Detection Isolation and Adaptation Principles Adaptation to noisy data Conclusions and Perspectives

3

• Ph. Bonnifait

Turn-by-turn navigation system

Map-matching and Route Planning GPS Navigation Map Destination Navigation function Driver interface for turn-by-turn guidance

4

• Ph. Bonnifait

Map-Aided ADAS

Example: Intersection Warning

Map-matching and Electronic Horizon computation GPS Navigation Map Navigation function Vehicle sensors CAN bus Speed Yaw rate Odometer ... EH Driver commands

 Electronic Horizon (EH): representation of oncoming context events

(e.g., curve, speed limits, intersection, etc.)

5

• Ph. Bonnifait

Map-Aided ADAS

Example: Intersection Warning

Map-matching and Electronic Horizon computation GPS Navigation Map Navigation function Vehicle sensors CAN bus Engine control Brakes Speed Yaw rate Odometer ... EH Driver commands Cluster / HMI Intersection warning Distance to intersection Current Speed Warning request Braking request

 Electronic Horizon (EH): representation of oncoming context events

(e.g., curve, speed limits, intersection, etc.)

6

• Ph. Bonnifait

Problem Statement

 Map errors may be due to:

• Errors during the mapping process.
• Evolution of road network.

 What happens if the map is wrong ?

• Uncomfortable and unsafe situations.
• Repetitive ADAS malfunctions.

7

• Ph. Bonnifait

Curve warning system

Navigation map

30 m

8

• Ph. Bonnifait

Curve warning system

GPS logs on top of the vehicle navigation map

30 m

9

• Ph. Bonnifait

Missed detection of the road bend.

Curve warning system

30 m

10

• Ph. Bonnifait

Problem Statement

• 1. Evaluate navigation system integrity in real-time.
• 2. Provide a correction when necessary.
• 3. Use only on board vehicle sensors.

11

• Ph. Bonnifait

Outline Context and Problem Statement Fault Detection Isolation and Adaptation

System architecture Methods for structural and geometrical faults Experimental results

Adaptation to noisy data Conclusions and Perspectives

12

• Ph. Bonnifait

Ability of the vehicle to assess the confidence associated to navigation information using redundant information from on board sensors.  every trip on the same road adds redundancy To provide a reliable confidence indicator to avoid client systems malfunctions.

Autonomous integrity monitoring

Map Matching Client Systems GNSS Navigation Function Navigation Map EH

13

• Ph. Bonnifait

Architecture for Autonomous Integrity Monitoring

Map Matching Client Systems Knowledge of fault GNSS2 Proprioceptive sensors GNSS1 Navigation Function Smart front camera Correction Integrity of Navigation Information Memory Navigation Map EH Don’t use Unknown Use

14

• Ph. Bonnifait

Definitions

Fault: Error generative process. Error: Discrepancy between measured value and true value. Failure: Time when a function exceeds the acceptable value.

A1 A2

A

15

• Ph. Bonnifait

Map Matching Client Systems Knowledge of fault GNSS Proprioceptive sensors GNSS Navigation Function Smart front camera Correction Integrity of Navigation Information Memory Navigation Map EH Don’t use Unknown Use

Case of Navigation

Fault: GNSS multipath; Wrong road candidate selected by map-matching ; Wrong representation of the road network. Error: Discrepancy between value in the EH and true value. Failure: Dysfunction of a client ADAS or autonomous driving function.

16

• Ph. Bonnifait

Map Geometric Faults

17

• Ph. Bonnifait

Geometric Fault Detection, Isolation and Adaptation

The vehicle position is encoded with:

The curvilinear abscissa s The trip number k

18

• Ph. Bonnifait

Geometric Fault Detection, Isolation and Adaptation

The vehicle position is encoded with:

The curvilinear abscissa s The trip number k

19

• Ph. Bonnifait

Geometric Fault Detection, Isolation and Adaptation

The vehicle position is encoded with:

The curvilinear abscissa s The trip number k

FDIA is based on the comparison of vehicle position estimates:

G from vehicle sensors N from the Navigation function estimate

20

• Ph. Bonnifait

Geometric fault detection, isolation and adaptation

Detection: Determine whether an estimate is affected by a fault Isolation: Determine which estimate is affected by a fault Adaptation: Identify a fault free estimate to provide it to client systems

21

• Ph. Bonnifait

Assumptions

When travelling several times on a road, the vehicle follows the same path with small deviations At a given abscissa:

• Faulty vehicle position estimates from sensors are different

from one trip to the other.

• Faulty vehicle position estimates from the navigation are

always the same.

• Faults on the vehicle position estimates from sensors and from

the navigation are different from each other.

22

• Ph. Bonnifait

Method

 First vehicle trip  Two independent estimates of the vehicle position:

• G1 (from vehicle sensors)
• N1 (from navigation system)

 Observed residual:  G1 affected by a fault:  N1 affected by a fault:

23

• Ph. Bonnifait

Possible outcomes

G1 = N1 Both estimates are fault-free and

Faults on estimates from sensors and from the navigation are assumed to be different from each other The residual is therefore the result of a Boolean OR:

• G1 ≠ N1



One estimate is faulty and and



Both estimates are faulty

and

Faults and residuals

24

• Ph. Bonnifait

Compute the residual based on the available estimates Find this residual in the truth table Provide the knowledge of fault to client systems

Method

25

• Ph. Bonnifait

Illustrative example: First trip

G1 ≠ N1 Both estimates are possibly faulty A fault is detected but not isolated The method returns Unknown

Abscissa s (m) Abscissa s (m)

s = 10m

26

• Ph. Bonnifait

Illustrative example: First trip

G1 = N1 There is no fault The method returns Use

Abscissa s (m) Abscissa s (m)

s = 20m

27

• Ph. Bonnifait

Illustrative example: First trip

G1 ≠ N1 Both estimates are possibly faulty A fault is detected but not isolated The method returns Unknown

Abscissa s (m) Abscissa s (m)

s = 40m

28

• Ph. Bonnifait

 Second vehicle trip  Two new estimates of the vehicle position at the same abscissa

• G2 (from vehicle sensors)
• N2 (from navigation system)

N1 G1 N2 G2

 Observed residual vector:

Using several trips

29

• Ph. Bonnifait

Possible outcomes when comparing two estimates from sensors G1 and G2

G1 = G2

Both estimates are fault-free and

Errors on estimates from sensors are assumed to be different from one trip to the

• ther

The residual is therefore the result of a Boolean OR:

• G1 ≠ G2

 One estimate is faulty and and  Both estimates are faulty and

Faults and residuals

30

• Ph. Bonnifait

Possible outcomes when comparing two estimates from Navigation N1 and N2

N1 = N2 : Map faults

Both estimates are fault-free and Both estimates are faulty and

Errors on the vehicle position estimates from the navigation are always the same The residual is therefore the result of a Boolean Exclusive OR:

• N1 ≠ N2 Matching faults

 One estimate is faulty and and

Faults and residuals

31

• Ph. Bonnifait

Truth table for two trips

32

• Ph. Bonnifait

Illustrative example: Second trip

This residual is unique in the table, isolation is done The current navigation is found not faulty The output Use is provided to client systems

s (m) s (m)

s = 10m

33

• Ph. Bonnifait

Illustrative example: Second trip

This residual is unique in the table, isolation is done The current navigation is found not faulty The output Use is provided to client systems

s (m) s (m)

s = 20m

34

• Ph. Bonnifait

Illustrative example: Second trip

This residual is four times in the table, fault is detected but not isolated The output Unknown is provided to client systems

s (m) s (m)

s = 40m

35

• Ph. Bonnifait

 Guaranteed detection of

faults

Formalism properties

36

• Ph. Bonnifait

 Guaranteed detection of

faults

 Conservation of residual

isolability

Formalism properties

37

• Ph. Bonnifait

 Guaranteed detection of

faults

 Conservation of residual

isolability

 Isolation convergence

Ratio of adverse residuals

• One trip: q(1) = 3/4
• Two trips: q(2) = 6/16=3/8
• Infinity: q(Inf)  0

Formalism properties

38

• Ph. Bonnifait

 Guaranteed detection of

faults

 Conservation of residual

isolability

 Isolation convergence

Ratio of adverse residuals

• One trip: q(1) = 3/4
• Two trips: q(2) = 3/8
• Infinity: q(Inf)  0

 Adaptation

There is at least one non faulty estimate in isolable sets of faults

Formalism properties

39

• Ph. Bonnifait

 Guaranteed detection of

faults

 Conservation of residual

isolability

 Isolation convergence

Ratio of adverse residuals

• One trip: q(1) = 3/4
• Two trips: q(2) = 3/8
• Infinity: q(Inf)  0

 Adaptation

There is at least one non faulty estimate in isolable sets of faults

 Conservation of adaptation

When isolation is performed, there will be at least one non faulty estimate at the next trip. This will make Adaptation possible

Formalism properties

40

• Ph. Bonnifait

Illustrative example: Third trip

s (m)

41

• Ph. Bonnifait

FDIA Algorithm for on board implementation

Before hand computation of the truth tables In real-time:

42

• Ph. Bonnifait

Experimental validation

Probe Vehicle Software tools

u-blox GPS receivers CAN-Bus

• Vehicle speed
• Wheel Speed
• Yaw rate
• Odometer...

GPS + IMU as ground truth for localization Real-time data acquisition Data replay OSM Navigation map Electronic Horizon generation Fault Detection, Isolation and Adaptation GPS N, road id, s G

43

• Ph. Bonnifait

True Validations (TV):

Correct navigation point identified as not-faulty

True Isolations (TI):

Faulty navigation point isolated by the method

Overall efficiency rate (OER): Information availability rate (IAR):

Metrics

44

• Ph. Bonnifait
• used navigation map in yellow
• correct map is in grey in background
• vehicle goes from left to right
• first trip is in blue; second trip in purple

Rural results

500 m

New bridge New road Parallel road

45

• Ph. Bonnifait

Rural results

First Trip

Overall Efficiency Rate = 100% Information Availability Rate = 77%

Second Trip

Overall Efficiency Rate = 100% Information Availability Rate = 100%

Good performance

Real map geometric faults Simple GPS conditions

46

• Ph. Bonnifait

Outline Context and Problem Statement Fault Detection Isolation and Adaptation Extension to Handle Uncertainties

Page’s trend test Integration into the FDIA method Experimental results

Conclusions and Perspectives

47

• Ph. Bonnifait

Noisy Position Estimates

Deterministic FDIA method Noise on position estimate particularly in urban environment Proposed solution

Statistical analysis of GPS and Navigation estimates

48

• Ph. Bonnifait

System Architecture

Map Matching Client ADAS Vehicle position estimation Fault Detection, Isolation and Adaptation G N Knowledge

• f fault

GNSS Proprioceptive sensors GNSS Navigation Navigation Map Smart front camera Correction EH Page’s trend test N=G ?

49

• Ph. Bonnifait

Page’s trend test

Detection of a change in the mean of a random variable Hypotheses

di: distance between estimates μ0 and μ1: mean of d before and after the change in the mean r: time of the change in the mean b: noise.

Sequential likelihood ratio testing Page’s trend test localizes the mean change with a minimized delay

50

• Ph. Bonnifait

Example

51

• Ph. Bonnifait

Example

52

• Ph. Bonnifait

Example

53

• Ph. Bonnifait

Example

54

• Ph. Bonnifait

Example

55

• Ph. Bonnifait

Example

56

• Ph. Bonnifait

Example

57

• Ph. Bonnifait

Implementation in the FDIA framework

Page’s test provides the value of rGN When the decision variable is greater than 0 and lower than the threshold, the FDIA is delayed

58

• Ph. Bonnifait

Implementation in the FDIA framework

When Page’s trend test finally settles, FDIA is run at every buffered abscissa. Benefits of the test in this example:

False detection avoided at s = 11. False validation avoided at s = 15.

59

• Ph. Bonnifait

Random fault injection in maps

Purpose

Generates a variety of map faults Provides quantitative results

Principles

Add noise on position of road shape nodes Deletes some of the shape nodes

50 m 50 m

60

• Ph. Bonnifait

Urban Results

Five maps with random faults

Generated based on a high quality lane level map

Three trips clockwise

Purple lines

Three trips anticlockwise

Blue lines

50 m

61

• Ph. Bonnifait

Urban Results

Overall Efficiency

Evenly due to True Validations and True Isolations Gating and temporal data re-sampling effects

62

• Ph. Bonnifait

Urban Results

Information Availability

Low IAR at trip 1 due to large proportion of faulty map areas and faulty GPS faults Isolation convergence property verified

63

• Ph. Bonnifait

Urban Results

With Page’s test:

False Validation rate decreases.

Less faulty map points are identified as correct.  The output “use” is more reliable

False Isolation rate increases.

More correct map points are identified as faulty.  More unjustified “don’t use”.

The FDIA method is more cautious but client systems are deactivated more frequently.

64

• Ph. Bonnifait

Conclusion Context

Map-aided ADAS Constant evolution of the road network Black box systems in passenger vehicles

Contributions

An integrity monitoring architecture that uses repetitive trips A framework for geometric fault detection, isolation and adaptation An extension of the framework with Page’s test Tests and evaluations on real vehicle

65

• Ph. Bonnifait

Thank you for your attention!

Associated Publications:

• 1. C. Zinoune, Ph. Bonnifait, and J. Ibanez-Guzman. “Detection of missing roundabouts in

maps for driving assistance systems”. In Intelligent Vehicles Symposium, IEEE, 2012.

• 2. C. Zinoune, Ph. Bonnifait, and J. Ibanez-Guzman. “A sequential test for autonomous

localisation of map errors for driving assistance systems”. In Intelligent Transportation Systems (ITSC), 2012 15th International IEEE Conference on, pages 1377–1382, 2012.

• 3. C. Zinoune, Ph. Bonnifait, and J. Ibanez-Guzman. “Integrity Monitoring of Navigation

Systems using Repetitive Journeys. In Intelligent Vehicles Symposium”, IEEE, 2014.

• 4. C. Zinoune, Ph. Bonnifait, and J. Ibanez-Guzman. “Sequential FDIA for Autonomous

Integrity Monitoring of Navigation Maps on board Vehicles” to appear in IEEE Transactions on Intelligent Transportation Systems

Patent:

• C. Zinoune, Ph. Bonnifait, and J. Ibanez-Guzman. Process of detection of roundabouts for an

application conveys. INPI Patent Number FR2997183.