Wireless Sensor Network based Shooter Localization

Transcription

1 Wireless Sensor Network based Shooter Localization Miklos Maroti, Akos Ledeczi, Gyula Simon, Gyorgy Balogh, Branislav Kusy, Andras Nadas, Gabor Pap, Janos Sallai ISIS - Vanderbilt University

2 Overview CONOPS Support: Fast and accurate enemy shooter localization is key in reducing friendly casualties and neutralizing enemy combatants Different approach compared to other acoustic shooter localization systems (e.g., BulletEars, Pilar, Ferret, Boomerang) Ad-hoc wireless network of cheap acoustic sensors is used to accurately locate enemy shooters in urban environment Challenges Severely resource constrained nodes and communication bandwidth Extreme multipath effects in urban environment Simultaneous shot resolution Performance Average accuracy: 1 meter Latency: 2 seconds Simultaneous shot resolution: 6 shots per second

3 Technical Approach Detect TOA of acoustic shockwave and muzzle blast MICA2 mote (UC Berkeley and Crossbow) Acoustic sensor board (Vanderbilt): Three acoustic channels (only one is used) High-speed AD converters FPGA for signal processing: shockwave and muzzle blast detection on board Timestamp of shockwave and/or muzzle blast is sent to mote Motes route time of arrival data to base station Base station fuses data, estimates shooter position and displays result

4 Technical Approach: System Architecture Sensorboard Time Sync User Interface Muzzle Blast & Shockwave Detector Sensorboard Interface Data Encoder Time Sync Sensor Fusion Sensor Location Sensorboard Config/ Monitor Data Recorder Message Routing Message Center Remote Controller Stack Monitor Remote Control Download Manager Plotter Logger SENSORBOARD MICA2 MOTE BASE STATION I 2 C UART

5 Acoustic Sensor Board: False Alarm Rejection Speech Fragment Black Hawk Helicopter Clap Shockwave and Muzzle Blast 1ms/div

6 Acoustic Sensor Board: Different Weapon Types M16 FMG M16 Blank M16 SRTA MG1 Blank AK47 FMG Muzzle Blast 1ms/div Shockwave Muzzle Blast 1ms/div M16 FMG N-wave (zoom) 100µs/div

7 Acoustic Sensor Board Xilinx Spartan-II FPGA Three acoustic channels Only one is used in shooter localization Selection of the microphone Max 1 MHz sampling rate, 12-bit ADC ZC encoding Feature extraction Data compression

8 MICA2 hardware 8 MHz, 8-bit microcontroller 4 KB of RAM, 128 KB of ROM Wireless communication Packet is 29 bytes data, 7 bytes header 30 packets per second under no collisions feet range Power management 2-3 days of continuously operation on two AA batteries 2-3 of months when sleeping Various sensors and actuators boards Cost: $150 from XBow

9 Middleware Services: Time Synchronization Requirement sound travels one foot per millisecond time synch error in the whole network should be less than 1 millisecond (less than 1 ft error) Algorithm selected leader broadcasts its time receivers maintain a table of global-local time pairs receivers calculate clock offset and skew using linear regression receivers rebroadcast the global time Performance: ±1.5 µsec per hop error Low overhead: one timesynch round per minute (i.e., one message per minute per mote)

10 Time Synchronization Primitive: Time Stamping of Radio Messages sender receiver sync byte (~417 µs) handling delay (95% 0-1 µs) (5% 1-20 µs) header 5 hw and sw delay (~1386 µs) bit-offset (~0-365 µs) min min data time stamp 4 5 hw interrupt sw interrupt Mica2: 1.2 µs average error, 4.5 µs maximum error average Mica2Dot: 4 µs average, 12 µs maximum error

11 Time Synchronization: Experimental Evaluation 50 1 message per 30 seconds per mote % layout and links: average error (µs) maximum error (µs) syncronized motes (%) 90.00% 80.00% 70.00% m ic ro s e c o n ds :00 0:10 0:20 0:30 0:40 0:50 1:00 1:10 1:20 1:30 1:40 1:50 2:00 2:10 Time (hh:mm) A. B. C. D. E % 50.00% 40.00% 30.00% 20.00% 10.00% 0.00% p e rce n tag e first leader second leader A. All motes are turned on B. The first leader is turned off C. Randomly selected motes were reset every 30 seconds D. Half of the motes were switched off E. All motes were switched back on

12 Middleware Services: Message Routing Requirement: acoustic event triggers many motes at once; all need to get their data to base station with low latency Approach: convergecast to root using our directed flood routing framework: Ad-hoc routing Automatic aggregation Implicit acknowledgments Performance: when max distance from root is 5 hops, base station receives ~15 measurements in the first second

13 Message Routing: Channel Behavior MICA2 under no load: single mote is transmitting Effective region (95% delivery rate) is 0-10 feet Transitional region (5-95% delivery rate) is feet MICA2 under heavy load: most motes transmit Effective region: 5 feet, transitional region: 5-30 feet 70% of the motes in the transitional region receive messages with less than 30% Polite (never transmits) and impolite (causes collision) motes Use probabilistic methods: rely on the unreliable It is more probable that one of the motes with less that 30% delivery rate will receive a broadcasted message than one with a higher rate Do not limit the next hop to a single node Long unreliable links can route messages faster than short reliable ones

14 Message Routing: Flooding Policies Each flooding policy defines a state machine that describe the life cycle of data packets On each node each data packet is in one of the states Actions: received, sent, aged States are numbered from 0 (initial state) to 255 (final state) Packets with low numbered states are more important Packets with even numbered states are eligible for transmission r 0 s 1 a 3 a 6 s a 5 a 7 9 r + s a a 253

15 Fat Spanning-Tree Convergecast A spanning tree is formed Each node needs to know the node ID of its parent, grandparent, great-grandparent, and great-great-grandparent The rank of the node is the node ID of its grandparent If the rank of the sender is the node ID of the grandparent of the receiver, then the sender is at the same distance the node ID of the receiver or its parent, then the sender is further from the root than the receiver the node ID of the great- or great-great-grand parent of the receiver, then the sender is closer to the root non of the above: not in the same channel or further away. Increases the reliability and robustness of tree routing protocols Scales linearly with distance gradient flooding fat spanning-tree flooding

16 Middleware: Self Localization

17 Sensor Fusion: Consistency Surface t 1 t 2 d 2 d 1? f(x,y) d 3 t 3 d 4 t 4 f(x,y) = #datapoints in window t 2 -d 2 /v outlier t 3 -d 3 /v t 4 -d 4 /v t 1 -d 1 /v time

18 Sensor Fusion: Multiple Shots Find the global maximum of the consistency surface If the global maximum is above a threshold then this position is the 1 st estimated shooter position Remove the corresponding measurements from the data set Recalculate surface/ global maximum (next shots) Shot and its detected consistent echo have the same shot time echo elimination 1 st shot 2 nd shot echo 2 nd shot

19 System Performance Latency: 2 seconds Average accuracy in 2D (x,y): 0.64 meter; in 3D (x,y,z): 1.5 meter Results below based on: 71 single SRTA shots; 20 different positions (McKenna MOUT Site, Ft. Benning); 60-mote network; 100x40 m area Experiment Results (SRTA) Representative Distribution of 2D Location Data 47% 35% 1m Experiment Results (SRTA) Representative Distribution of 3D Location Data 17% 24% 1m % 2m 41% 2m 3% 3m 1 17% 3m % 5m Location error in meters Location error in meters

20 Observed Hardware Failures Noisy or non-functional ADC (3%) Memory corruption (only one mote) One-way radio, only receive or transmit (3%) Battery exploded (only one, left in the programming board) Corrupted fuses (5%) Non-functional LEDs (1%) Motes overrun by cars and people (2%) Under direct sunlight the ADC and CPU clock of the motes changed their characteristics

21 Development Tools

22 Unionized NEST Workers

23 Middleware Development in Tennessee

24 Oops, we have to recruit a new graduate student again