On-the-fly Route Planning for Mono-UAV Surveillance Missions M. - - PowerPoint PPT Presentation

• M. Soulignac1 F. Gaillard1 C. Dinont1 G. Marchalot2

On-the-fly Route Planning for Mono-UAV Surveillance Missions

UK PlanSIG'13 Edinburgh, 29 January 2014

1 ISEN-Lille

2 THALES Airborne Systems

Mono-UAV Surveillance Mission Planning

Soulignac et al. UK PlanSIG'13 Edinburgh, 29 January 2014

3

A Mono-UAV surveillance mission

UAV Detected target Identified target Non-detected target Identification range Detection range Targets (unknown position and route) Maximal targets velocity known a priori

4 End of the mission pD near 100% and pI near 100%

Mission Planning

Expected output

Beginning of the mission pD : % of detected targets pI : % of identified targets

5

Our approach

Projection on Pattern (POP)

Sweeping Pattern Local deviations => pD high => maximize pI and maintain pD high

UAV route

The Flight Pattern

Soulignac et al. UK PlanSIG'13 Edinburgh, 29 January 2014

7

(For static targets)

Basic Sweeping Pattern

forward backward

rD

area to cover

H

flight pattern repetitions of the pattern

8

Non-detected target

(For static targets)

Basic Sweeping Pattern

forward backward

T

9

(For static targets)

Basic Sweeping Pattern

Detected target forward backward A target missed during the forward move is ensured be detected during the backward move

T

10

Enhanced Sweeping Pattern

(For moving targets)

Non-detected target

vT T

forward backward

11

(For moving targets)

Non-detected target

Enhanced Sweeping Pattern

vT T

forward backward

12

with

(For moving targets)

Detected target

the velocity ratio between UAV and targets

Enhanced Sweeping Pattern dT = f(αv )

(see our paper for the expression of f and a sketch of proof)

Pattern narrowing

T

forward backward

13

Back to our example

= 180 knots = 7 knots

Situation at the beginning of the mission

Critical targets

14

Back to our example

Situation after executing one instance of the pattern

All the targets have been detected

Critical targets

(and the targets Ti at a distance to the pattern have been identified)

POP : Projection On Pattern

Soulignac et al. UK PlanSIG'13 Edinburgh, 29 January 2014

16

Route Planning Problem

How to investigate T1 and T2 while following the flight pattern ?

T1 T2

17

Targets projection

Target T1

T1 T2

61.1 230.4 422.4 37.5 138.8 30.5

Arc length on the pattern Distance to the pattern Projected target P11 P12 P13

18

Targets projection

Target T1

T1 T2

61.1 230.4 422.4 37.5 138.8 30.5

P1

19

Targets projection

Target T2

T1 T2

15.2 431.3 251.9 50.4 149.5 51.9

P23 P22 P21

20

Targets projection

Target T2

T1 T2

15.2 431.3 251.9 50.4 149.5 51.9

P2

21

Targets insertion

T1 T2

431.3 61.1 200.0 280.0 480.6 42.9

P1 W1 W2 W3

W1

W2 W3 200.0 280.0 ... 480.6

Flight pattern (copy)

P1

P2 61.1 431.3

Projected targets Arc length Name increasing order of arc length P2

22

Targets insertion

T1 T2

431.3 61.1 200.0 280.0 480.6 42.9

P2 P1 W1 W2 W3

W1

W2 W3 200.0 280.0 ... 480.6

Flight pattern (copy)

P1

P2 61.1 431.3

Projected targets Arc length Name increasing order of arc length insertion preserving

• rder

23

Targets insertion

T1 T2

431.3 61.1 200.0 280.0 42.9

W1

W2 W3 200.0 280.0 480.6

Updated UAV route

P1

P2 61.1 431.3

P1 W1 W2 W3

480.6

P2

24

Targets insertion

T1 T2

431.3 61.1 200.0 280.0 42.9

W1

W2 W3 200.0 280.0 480.6

Updated UAV route

P1

P2 61.1 431.3

P1 W1 W2 W3

480.6

P2 Postponed identification task

25

Result on an entire mission

Beginning of the mission pD = 100 % pI = 99.1 % End of the mission

POP

26

Result on an entire mission

End of the mission The UAV route can be improved 3 enhancements proposed

Enhancements

Soulignac et al. UK PlanSIG'13 Edinburgh, 29 January 2014

28

Target move anticipation

Without

T

T is projected on the closest part of the pattern, regardless

• f its heading.

T

target projection resulting flight plan

29

Target move anticipation

A B

Elapsed time from A to B : 9753 s.

= 180 knots = 6 knots

Without

Executed trajectory to identify T : T

30

Target move anticipation

With

Estimated target position at interception time1

1 assuming a rectilinear motion of T

T

Estimated UAV position at interception time2

2 assuming a strict pattern following

target projection resulting flight plan

31

Target move anticipation

A B

Elapsed time from A to B : 9638 s (previously 9753) Cost reduction : 1%

= 180 knots = 6 knots

T Executed trajectory to identify T :

With

32

Amplified Pattern Narrowing

Without

Missed target

Repeated deviations ⇒ UAV delay ⇒ Missed targets

33

Amplified Pattern Narrowing

Detected target (previously missed) No narrowing Amplified narrowing

With

Normal narrowing

34

Amplified Pattern Narrowing

Detected target (previously missed) No narrowing Amplified narrowing

With

dT = f(α 'v )

Normal narrowing

α 'v = g(α v , D) ; g adjusts α v according to the UAV delay D

(see our paper for the expression of g)

35

Local TSP

Without Wnext

Distance to Wnext : 296 NM Identifying targets by increasing arc length ⇒ oscillations

36

Local TSP

Distance to Wnext : 296 NM TSP tour to improve

Wnext Without

Identifying targets by increasing arc length ⇒ oscillations

37

Local TSP

Distance to Wnext : 229 NM (previously 296)

Cost reduction : 23%

Wnext

Tour improved by 2-opt

With

Demo

Soulignac et al. UK PlanSIG'13 Edinburgh, 29 January 2014

39

Demo

POP within the multi-agent platform APM(Robot)

Simulation results

Soulignac et al. UK PlanSIG'13 Edinburgh, 29 January 2014

41

Experimental protocol

Targets density

High density Low density L = 200 NM H = 200 NM

about 40 targets about 120 targets

42

Experimental protocol

Parameter variants

43

Computation time

Sensibility to density

High density Low density

44

Computation time

Sensibility to density

Computation time under 10 ms in 99.9%

• f simulations

Computation time under 10 ms in 98.1%

• f simulations

Parameter variants Computation time mostly under 10 ms even with :

• a high density of

targets

• all enhancements

enabled

45

Highest values for pD + pI

Low density High density

And corresponding parameter variants

46

pD + pI > 190% pD + pI < 190%

Low density High density

rI (NM) vT (knots) vT (knots)

Values of vT and rI leading to pD + pI > 190%

Conclusions and perspectives

Soulignac et al. UK PlanSIG'13 Edinburgh, 29 January 2014

48

Presented work

POP (Projection On Pattern) = Enhanced (i.e. narrowed) sweeping pattern + Target projections on this pattern + Target identification tasks ordered by increasing values

• f arc length

3 enhancements :

• target move anticipation
• amplified narrowing
• local TSP

49

Conclusions of simulation results

The best combination of enhancements depends on the

• context. Enabling all enhancements does not necessarily

lead to the best performances. High detection and identification performances can be

• btained for reasonable targets speeds and identification

range Computation time is mostly under 10ms, allowing on-the- fly replanning

50

Ongoing works

• Extension to multiple

cognitive UAVs

• Encouraging results

with mTSP algorithms

• Dynamic identification

tasks allocation

51

End of talk

Thanks for your attention. Slides available on my webpage.

Appendices

Soulignac et al. UK PlanSIG'13 Edinburgh, 29 January 2014

Enhanced Sweeping Pattern

Soulignac et al. UK PlanSIG'13 Edinburgh, 29 January 2014

54

Computation of dT

Enhanced Sweeping Pattern

Detected target

55

with

Enhanced Sweeping Pattern

Computation of dT

Detected target

56

Enhanced Sweeping Pattern

Root 1 Root 2 Root 1 Root 2 Downward move Upward move Keep the smallest root

Enhanced Sweeping Pattern

Computation of dT

57

Lower bound for

Enhanced Sweeping Pattern

The UAV must be faster than the targets

Maximal narrowing No narrowing (basic pattern)

Amplified Pattern Narrowing

Soulignac et al. UK PlanSIG'13 Edinburgh, 29 January 2014

59

Amplified Pattern Narrowing

Principle

• vercost

Arrival at T Estimated arrival at T' = T +

• vercost

vUAV Situation 1 : Strict pattern following Situation 2 : Pattern following + local deviations

The situation 2 is equivalent to the situation 1 with a UAV speed v'UAV < vUAV such that :

l1 l2 l3

v'UAV = l1 + l2 + l3 T'

60

Principle

Amplified Pattern Narrowing

simulates a slower UAV relatively to targets (or faster targets relatively to the UAV). Amplified pattern narrowing consists in using the velocity ratio : instead of in the pattern narrowing process.

estimated end of the mission with deviations estimated end of the mission with deviations

( dT is computed by dT = f(α '

v ) )

Local TSP

Soulignac et al. UK PlanSIG'13 Edinburgh, 29 January 2014

62

Local TSP

Tour improver Small TSP tours (about 10 cities in average for a high density of targets) The choice of the tour improver has no significant impact on the solution quality Our choice : 2-opt

• Excellent response time (to be seen in the simulation results)

* measured on 10000 10-cities TSP tours. The optimal solutions have been computed using branch and bound with the Held-Karp relaxation.

• Solution at 0.007% of optimal*
• Easy to implement and customize

63

Local TSP

2-opt 2-opt improves a given tour by performing 2-edge flips. The resulting tour is guaranteed to be non-self-crossing.

Nearest Neighbor Tour

64

Local TSP

Step 1

2-opt improves a given tour by performing 2-edge flips. The resulting tour is guaranteed to be non-self-crossing. 2-opt

65

Local TSP

Step 2

2-opt improves a given tour by performing 2-edge flips. The resulting tour is guaranteed to be non-self-crossing. 2-opt

66

Local TSP

Step 9 ( ... )

2-opt improves a given tour by performing 2-edge flips. The resulting tour is guaranteed to be non-self-crossing. 2-opt

67

Local TSP

Step 10

2-opt improves a given tour by performing 2-edge flips. The resulting tour is guaranteed to be non-self-crossing. 2-opt

68

Target move anticipation

Interpretation

Projection without anticipation (projection on pattern) Projection with anticipation (spatio-temporal projection) Pattern waypoint

Simulation results

Soulignac et al. UK PlanSIG'13 Edinburgh, 29 January 2014

70

Experimental protocol

Context : targets speed & identification range

Targets speed vT (knots) vT / vUAV vUAV = 180 knots 4.5 5.14 6.0 7.2 9.0 12.0 18.0 10% 6.6% 5% 4% 3.3% 2.8% 2.5% vT / vUAV vUAV = 180 knots rI (NM) rI / rD 5 10 15 20 25 30 35 100% 85.7% 71.4% 57.1% 42.8% 28.5% 14.2% Identification range rD = 35 NM

71

Experimental protocol

Measures

For each 4-uplet (density, vT , rI, parameter variant), we measured :

• n the same 30 randomly generated environments, and computed some

statistics (mean, standard deviation, percentiles, etc.) The 13230 simulations required 10.9 CPU-days with a low density context and 18.9 CPU-days with a high density context.

• the computation time
• the sum pD + pI (a way to check that both pD and pI are high)

72

Computation time

Sensibility to parameters

No rerouting Rerouting without local TSP Rerouting with local TSP

High density Low density

75th 50th 25th percentiles