Shooter Localization with Wireless Sensor Networks
description
Transcript of Shooter Localization with Wireless Sensor Networks
Institute for Software Integrated SystemsVanderbilt University
Shooter Localizationwith
Wireless Sensor Networks
Akos LedecziAssociate Professor
Copyright © 2004-2011, Vanderbilt University2
Evolution Many single-channel acoustic sensors
2003-2005 Designed for urban operation:
Multipath elimination Multiple simultaneous shot resolution
1-meter 3D accuracy within network No classification DARPA NEST Program
Few multi-channel acoustic sensors 2005-2006 Helmet-mounted, 4-channel acoustic sensor node Single sensor operation: localization Networked operation: trajectory and caliber estimation and weapon classification 1-degree bearing accuracy DARPA ASSIST Program
Few single-channel acoustic sensors 2010- Mobile phone-based system Single sensor operation: miss distance and range estimation* Networked operation: trajectory and caliber estimation and weapon classification DARPA Transformative Applications Program
Copyright © 2004-2011, Vanderbilt University3
Wireless Sensor Network-BasedCountersniper System
Copyright © 2004-2011, Vanderbilt University4
RITS: Routing Integrated Time Synch
reactive protocol, synchronizes after the event was registered (post-facto)
maintains the age of event instead of the global time and computes the local time of event at the data fusion node
power efficient, virtually no communication overhead, can be highly accurate
Δt1 + Δt2 + Δt3 Troot
Tevent = Troot - Δt1 - Δt2 - Δt3
node1time
node2time
node3time
sinktime
Tevent
Δt1
Δt2
Δt3
event
sink
average time synchronization error histogram
0%
5%
10%
15%
20%
25%
30%
0 2 3 5 6 8 10 11 13
synchronization error (microseconds)
perc
enta
ge
- ~50 node experiment- 4.4 μs average error, 74 μs maximum error
in the test of 200 rounds
Copyright © 2004-2011, Vanderbilt University 5
t2
t1
t4
t3
d1
f(x,y)?
d3
d4
d2
Shot #1 @ (x1,y1,T1)
Shot #2 @ (x2,y2,T2)
Echo #1 @ (x3,y3,T1)
timet2 – d2/vt3 – d3/v
t1 – d1/vt4 – d4/v
f(x,y) = [max number of ticks in window] = 3Shot time estimate T
3 0 1
sliding window
Sensor Fusion
Copyright © 2004-2011, Vanderbilt University6
Experiments at McKenna MOUT site at Ft. Benning
NORTH
B1Church
Sep 2003: Baseline system Apr 2004: Multishot resolution
60 motes covered a 100x40m area Network diameter: ~7 hops Used blanks and Short Range Training
Ammunition (SRTA) Hundreds of shots fired from ~40 different
locations Single shooter, operating in semiautomatic
and burst mode in 2003 Up to four shooters and up to 10 shots per
second in 2004 M-16, M-4, no sniper rifle Variety of shooter locations (bell tower,
inside buildings/windows, behind mailbox, behind car, …) chosen to absorb acoustic energy, have limited line of sight on sensor networks
1 meter average 3D accuracy (0.6m in 2D)
Hand placed motes on surveyed points (sensor localization accuracy: ~ 0.3m)
Copyright © 2004-2011, Vanderbilt University 7
2.5D Display, Single shotRed circle: Shooter position
White dot: Sensor node
Small blue dot: Sensor Node that
detected current shot
Cyan circle: Sensor Node whose
data was used in localization
Yellow Area: Uncertainty
Copyright © 2004-2011, Vanderbilt University 8
2.5D Display, Multiple ShotsRed circle: Shooter position
White dot: Sensor node
Small blue dot: Sensor Node that
detected current shot
Cyan circle: Sensor Node whose
data was used in localization
Yellow Area: Uncertainty
Copyright © 2004-2011, Vanderbilt University9
Shooter Localization
VIDEO
Copyright © 2004-2011, Vanderbilt University
Soldier-Wearable Shooter Localization System
DARPA IPTO ASSIST
Muzzle blast Shockwave
Zigbee&
Bluetooth
Microphones
3-axis compassOptional
laptop display
PDA displayZigbee
Bluetooth
Bluetooth
Zigbee
Copyright © 2004-2011, Vanderbilt University11
Acoustic Sensor Board
Detect TOA and AOA of ballistic shockwave and muzzle blast using a single board
Acoustic sensor board: 4 acoustic channels w/ high-speed AD
converters FPGA for signal processing 3-axis digital compass Bluetooth MicaZ connectivity
Copyright © 2004-2011, Vanderbilt University12
Software Architecture PC/PDA (Java/Ewe)
User interface Local/central sensor fusion Location information from
external GPS
Sensor Board (VHDL/assembly) Custom DSP IP cores (detection) Soft processor macros (digital
compass, debug & test interface)
Communication bridge Shared memory paradigm
Mote (nesC/TinyOS): Data sharing across nodes Time synchronization Application Configuration &
Management (from a central point)
Copyright © 2004-2011, Vanderbilt University
Single Sensor Results Independent evaluation by NIST at
Aberdeen in 2006 Localization rate for single sensors:
range < 150m: 42% Range < 80m: 61%
Percentage of shots not localized by at least one single sensor alone (range < 150m): 13%
Accuracy: 0.9 degree in azimuth 5 m in range
Blue dots: sensorsBlack squares: targetsBlack line: trajectory estimateBlack dot: shooter position estimateWhite arrows: single sensor shooter estimates
13
Copyright © 2004-2011, Vanderbilt University
Sensor Fusion Localization: Single sensor: simple
analytical formula to compute shooter location based on Time of Arrival (ToA) and Angle of Arrival (AoA) of both shockwave and muzzle blast.
Localization: Multi-sensor: all available detections are utilized in a multiresolution search of a discrete multi-dimensional consistency function. Consistency function specifies how many observations agree on a given point in space and time.
Online caliber estimation based on measured ballistic shockwave length and miss distance given by the computed trajectory estimate.
Online weapon classification based on estimated caliber and muzzle velocity that is computed using the projectile velocity over the sensor web and the estimated range.
14
Copyright © 2004-2011, Vanderbilt University
Multi-Sensor Results
Localization Results
Classification Results
Independent evaluation by NIST at Aberdeen in 2006
Shots between 50 and 300m w/ 6 different weapons (3 calibers)
Trajectory was highly accurate Big range error at >200m was due to
a bug in the muzzle blast detection Caliber estimation was almost
perfect (rates are relative to localized shots, not all shots).
Classification for 4 out of 6 six weapons were excellent
At longer ranges it started to degrade as it needs range estimate, i.e. muzzle blast detections
M4 and M249 was too similar to each other and the test was the first time the system encountered these weapons
Sensors located on surveyed points with small position error. Manual orientation and then automatic calibration used. No mobility. 15
Copyright © 2004-2011, Vanderbilt University16
Test in Georgia in 2009
VIDEO
Copyright © 2004-2011, Vanderbilt University
Motivation for New Approach
Traditional WSN approach: Many single channel sensors distributed in the
environment Too many nodes needed
Wearable sensor approach: Few multi-channel sensors Needs to track self-orientation: Hard!
17
What can be done with a few single-channel sensors?
Copyright © 2004-2011, Vanderbilt University
Accurate miss distance estimation using a single microphone (i.e. phone) by measuring the shockwave length. Estimated accuracy: 1-2m.
Accurate range estimation using a single microphone (i.e. phone) utilizing the miss distance and the TDOA of the shockwave and the muzzle blast: Estimated accuracy: 5%.
Novel consistency function-based sensor fusion technique enables localization of shooter with as few as 5 phones even in the presence of GPS and other errors.
Custom headset will provide better performance offloading the computationally intensive operations from the phone increasing battery life.
SOLOMON: Shooter Localization with Mobile Phones
DARPA Transformative Apps Program
Muzzle blast Shockwave
Phone Network
Phone Network
18
Copyright © 2004-2011, Vanderbilt University
Relation between shockwave length (N-wave duration in the time domain) and miss distance [Whitham52]:
Miss Distance Estimation in Standalone Operation
19
T: shockwave lengthM: Mach speed of the bulletb: miss distancec: speed of soundd: bullet caliberl: bullet length
Using 168 shockwave detections of AK-47 shots fired from 50 to 130m from sensors, with miss distances ranging from 0 to 28m, the average absolute miss distance error is 1m.
Miss distance can be computed from the shockwave length, with assumptions on the weapon (caliber, length and speed of bullet):
b: miss distanceT: shockwave lengthk: weapon coefficient
Copyright © 2004-2011, Vanderbilt University
Range can be calculated using the miss distance, a projectile speed and the TDOA of the shockwave and the muzzle blast.
Range Estimation in Standalone Operation
Phone
Shooter
SM:QM:P:SP:PM:α:
rangemiss distanceorigin of shockwave heard at Mat the speed of bulletat the speed of soundshockwave cone angle
20
Using 168 AK-47 shot detections from ranges between 50 and 130 m gathered at Aberdeen in 2006 the average range estimate has ~5% error.
Copyright © 2004-2011, Vanderbilt University
High quality application-specific microphone with higher maximum sound pressure and faster recovery (Knowles VEK-H-30108)
Higher sampling rate for better shockwave length and miss distance estimation Off-loading the signal processing algorithm from the phone using a low-power ARM-Cortex
microcontroller real-time signal processing with lower jitter and latency better performance/power ratio
Wired and/or wireless phone interface supporting any Android handset device Bluetooth interface with Android 2.0 and later Analog signaling on the headset audio interface using software modems on both sides
Integrated temperature sensor for more accurate speed of sound estimation
Custom Headset
21
Copyright © 2004-2011, Vanderbilt University
Networked Operation1. Multilateration: find an initial shooter position estimate using muzzle
blast TDOAs optional
2. Trajectory search: minimize an error function in a predefined search space Inputs:
shockwave TDOAs shockwave length
Optimized parameters: trajectory weapon coefficient
Side effects: Bullet speed is computed Miss distances are available
• Final shooter localization: constrained triangulation using range estimates
• Weapon classification using weapon coefficient and bullet speed22
No known weapon assumption.
Copyright © 2004-2011, Vanderbilt University
Miss distance is proportional to the fourth power of the shockwave length.
Miss distance is linearly related to weapon coefficient.
MSE of the n best miss distances is used as a metric for the trajectory (n=5 is good in practice)
Error Function: Miss distance consistency
Optimize the weapon coefficient for the trajectory What is the best weapon coefficient for the evaluated trajectory? How good is the match? Which trajectory has the best match?
.M1
.M2
.M3
.M1
.M2
.M3
.M1
.M2
.M3
Copyright © 2004-2011, Vanderbilt University
Error function: Cone angle consistency
Pairwise shockwave TDOA-based trajectory angle consistency Given a trajectory, the shockwave TDOA of two nodes can be used
to compute the shockwave cone angle. We compute the shockwave cone angle for all pairs of nodes, and
use the variance of the most consistent subset of size n as the metric (n=5 is good in practice).
24
Mi: microphone i positionBi: position of bullet when shockwave reaches microphone iQi: point on trajectory closest to
microphone ibi: miss distancec: speed of soundα: shockwave cone angleΔt: shockwave TDOA
The multiple of the miss distance-based and the cone angle-based consistency metric is minimized.
Copyright © 2004-2011, Vanderbilt University
Final Shooter Localization
25
Trajectory is known at this point Miss distances are also known Bullet speed is also known Range to each sensor can be estimated without the known
weapon assumption! Constrained trilateration using ranges and the known trajectory
Multilateration Trilateration Composite
Copyright © 2004-2011, Vanderbilt University
Classification Based on weapon coefficient and projectile speed, the bullet coefficient (caliber and length) is
estimated Based on bullet coefficient, range and speed, the muzzle velocity can be estimated (using an
approximate deceleration profile) Caliber and muzzle velocity is characteristic of rifles
26
T: shockwave lengthv: bullet speed (over network)M: Mach speed of the bulletb: miss distancec: speed of soundd: bullet caliberl: bullet length
12.70mm
7.62mm
5.56mm
M107
M240AK47
M4 & M249
M16
Copyright © 2004-2011, Vanderbilt University
Evaluation
27
Out of 108 shots, 107 trajectories could be computed. Average trajectory angle error is 0.1 degree, with standard deviation of 1.3 degrees. Absolute trajectory angle error is 0.8 degree.
Out of 108 shots, 104 shooter positions could be computed. Average position error is 2.96m, which is better than the 5.45m error with the previous, multi-channel system.
Copyright © 2004-2011, Vanderbilt University
Results from a single soldier’s POV
28
Average range error is 0.2m, with standard deviation of 3.3m. Average absolute range error is 2.3m.
Average individual bearing error is 0.75 degree.
Copyright © 2004-2011, Vanderbilt University
Questions?
More information:
Sallai, J., Ledeczi, A., Volgyesi, P.: “Acoustic Shooter Localization with a Minimal Number ofSingle-Channel Wireless Sensor Nodes” SenSys 2011
Volgyesi, P., Balogh, G., Nadas, A, Nash, C., Ledeczi, A.: “Shooter Localization and Weapon Classification with Soldier-Wearable Networked Sensors” MobiSys 2007
Ledeczi, A. et al.: “Countersniper System for Urban Warfare,” ACM Transactions on Sensor Networks, Vol. 1, No. 2, pp. 153-177, November, 2005
29