In-Pavement Wireless
Sensor Network for Vehicle
Classification
AUTHORS:RAVNEET BAJWA, RAM
RAJAGOPAL, PRAVIN VARAIYA
AND ROBERT KAVALER
PRESENTER: XIANGYI GU
Outline
 Motivation
 Introduction
 Description
 Communication Protocol Design
 Experiment Setup
 Performance
 Conclusion & Future Work
Outline
 Motivation
 Introduction
 Description
 Communication Protocol Design
 Experiment Setup
 Performance
 Conclusion & Future Work
Motivation
 Intrusive technologies
 Piezoelectric sensors, inductive loops (examples)
 High installation and maintenance costs
 Non-intrusive technologies
 Infrared, video imaging(examples)
 Sensitive to traffic and weather condition
 Propose an alternative system base on a WSN that is
both cost effective and insensitive to
environmental conditions
Transportation agencies collect vehicle classification information to plan highway
maintenance programs, evaluate highway usage and optimize the deloyment of
various resources on the road
 Motivation
 Introduction
 Description
 Communication Protocol Design
 Experiment Setup
 Performance
 Conclusion & Future Work
Problem Statement
 Cars, buses, three-axle single unit trucks, and five-
axle single trailer trucks (classifying vehicles)
 A vehicle travels in a traffic lane at some varying
speed and we wish to count the number of axles and
the spacing between each axle in an accurate manner
Proposed WSN System
 Vibration sensor (accelerometer) embedded in the road
 Calculate the axle spacings
 Vehicle detection sensor (magnetometers)
 Report the arrival and departure times of a vehicle
 Access point (AP)
 Send commands to sensors
 Log the incoming data
 First in-pavement, easy
to deploy, WSN based
system for counting axles
and axle spacing
Outline
 Motivation
 Introduction
 Description
 Communication Protocol Design
 Experiment Setup
 Performance
 Conclusion & Future Work
Wireless Vehicle Detection Sensor
 Measures the changes in magnetic field to infer the local
presence of a vehicle
 Synchronous Nanopower Protocol(SNP), a
TDMA based protocol

Last 10 years with a single 7200 mAhr battery
 Given the arrival times tai and taj at the two
sensors i and j, the speed v will be
v = dij / |taj – tai|
 Estimate the length(L) of the vehicle
L = v(tdj - taj)
Wireless Vibration Sensor
•Sample the analog output of an accelerometer and transmit the data via a
radio
•Designing a sensor that measures pavement vibrations for axle detection
have many unique challenges
Sensor needs to be insensitive
to the vehicles traveling
in the neighboring lanes
Sample fast enough to
capture the transient vibrations
Insensitive to the truck engine
and environmental noise
The sensor has to be
well coupled to the road way
and be resistant to
heavy vehicle traffic
Challenges
•Sensor resolution target
is 500ug
•Bandwidth 50Hz
•Sampling frequency 512
Hz( > 5 times Nyquist
Frequency)
-Power consumption increases
for higher sampling rates
Axle detection and counting
Given vehicle speed measurement and reliable
measurement from the wireless vibration sensor, we still
need to construct an axle detection algorithm that has
good performance
There are two important challenges in detecting
individual axles:
B
A
The vibration signals
from successive axles
tend to blend.
.
In wide highway
lanes, vehicles can
experience
significant wander
Sensor Design
Resolution:Selecting an accelerometer
 SD1221-005 has higher sensitivity and lower noise
density
 However, it consumes more than 20 times the current
than MS9002.D and has to be operated at higher
voltage
 Both devices achieved the aimed minimum resolution
of 500 ug

Select MS9002.D due to its low operating voltage and low
current consumption
Noise: Filters for mitigating sound noise
 Accelerometer is sensitive to sound
 MS9002.D behaves like a microphone under the
device’s bandwidth
 3rd order low-pass filter with cutoff frequency of 50
Hz is sufficiently aggressive to filter out most of the
sound in the audible spectrum
Casing
 Sound isolation
 Protect the electronics from
rain water and oil spill on the
road
Circuit Description
 2.5 V supply voltage
 Amplifier with gain 10
 The gain of 10 reduces the range of
the accelerometer to ≈±225mg
 This is necessary in order to ensure
that the quantization noise from the
ADC is less than the noise from the accelerometer

Otherwise, the resolution of the system will be limited by ADC noise
 The reduced range is still sufficient
 For heavy trucks  ± 200 mg
Outline
 Motivation
 Introduction
 Description
 Communication Protocol Design
 Experiment Setup
 Performance
 Conclusion & Future Work
Communication Protocol Design
 MAC Layer
TDMA based
 Time is divided into multiple frames with each frames
about 125 ms long
 Each frame is further divided into 64 time slots
 Slot 0 is used by AP to send clock synchronization
information and other commands to the sensors
 AP assigns every node unique time slots and a node ID to
communicate with it.

Application Layer
 Sync Application
AP sends sync packets on a periodic basis
 Sensor node listens to sync packets every 125 ms
 When the clock converges to steady state, then is listens
for a sync packet only once in 30 s
 Sync application is also used to send commands
 Set Mode, Reset, Set Timeslot, Set RF, Download
Firmware, Set ID

Application Layer
 Accelerometer Application
 Idle Mode: accelerometer and related circuitry are turned off
by disabling the voltage regulator

Once every 30 s, the microcontroller and the transceiver wake up
and acquire the sync packet
Application Layer

Raw Data Mode: microcontroller wake up every 1/512 s, and
samples the analog output from accelerometer
32 samples at a sampling freq. 512Hz, and each sample containing
12 bits of information
 In every frame(125ms) we accumulate 96 bytes of information to
transmit
 To have a reasonable packet size, we fragment the data in two
parts, 48 bytes each, and transmit it using two different time slots
62.5ms apart

Application Layer
 Download Firmware Application
Reprogram the entire flash memory of a sensor node over
the air
 AP transmits new code repeatedly and the node updating
its code in small pieces
 Only the data that do not overwrite the current running
program are updated by the node

Axle Detection(ADET) Algorithm
Results of ADET on truck49( two single axles and one tandem axle), a(n) is the
measured acceleration in mg, e(n) is the scaled energy in mg2, and s(n) is smooth
energy in mg2. The red asterisks on s(n) are the axle locations found by ADET. By
reducing the minimum axle sepatation, the individual axles in the tandem axle can
also be detected as shown by black circle
Axle Detection(ADET) Algorithm
 Using data from 4 trucks at different speeds, we
observed the bandwidth of the energy signal and
empirically defined by M(v) = 900/v
 Low-pass filter is optional
 Minimum time separation ζ(v) was chosen by
assuming that the axles are at least 6ft apart
Wide Lane ADET Algorithm
 Wander movement in a lane
 Combining vibration readings from multiple sensors
 Delay  Di = di / v sennor2 will measures the peak enery a little later than sensor 1,
so the individual energy measurements nedd to be appropriately delayed
System representation of the adjustment made to correct vehicle wander.
The energy if the total signal at time n is the maximum of the energy of
the individual signals ei(n) is each sensor i.
Outline
 Motivation
 Introduction
 Description
 Communication Protocol Design
 Experiment Setup
 Performance
 Conclusion & Future Work
Experiment Setup
 4 vibration sensors and 4 vehicle detection sensor were
installed on California Highway I-680
 Vehicles come from Sunol Weigh
Station
 Slow down at weigh station

Easy to collect ground truth
 Data from 53 different trucks, ranging
from pickup trucks to 5-axle
commercial trucks
The WSN setup at Suno; site. DHMN are vehicle
detection sensors shereas I,J,K,L are vibration
sensors
Installation
 Boring a 4-inch diameter hole approximately 2.25 inches
deep
 Installed on a road in less than 20 minutes
 Installation of a small sensor is much cheaper and
convenient than installing special material pavements
required for piezoelectric sensors
Deployment Challenges
 Packet Drops

Drop rate was low(1%) (compare 50 feet away)
retransmit packets with a delay of 1 packet  drop rate is
almost 0

Packet 1, 2, 1again, 2again
 Vehicle Wander
Because vehicles are not taveling stright in a lane we would like to
choose data from vehicle sensor that was closest to the tires
 use Wide Lane ADET algorithm
 Sensor failure
 Sensor k did not work
 Vibration data was available from 3 sensors
Outline
 Motivation
 Introduction
 Description
 Communication Protocol Design
 Experiment Setup
 Performance
 Conclusion & Future Work
Vibration Sensor Performance
 Noise with no vehicle in vicinity
 414 ug RMS
 Truck was parked on top of the sensor with engine
were on vs. truck blew its horn

7% vs. 4%
 With a heavy truck traveled in the closed lane
 Sensor did not register any noticeable peaks
Axle Count
 Error  difference between the ground truth axle count
and the estimated axle count
 By combining the measurements from all sensors, the
algorithm always gives the correct axle count
 Error results form the wander movement( Strongly affected
by truck wander
Performance of ADET using individual sensors and combinations of sensors.
Count Err. Is the difference between the ground truth and ADET estimate.
Under each sensor column is the observed frequency of the errors.
Axle Spacing
 Left: for tandem axle
 Middle: pick up trucks, small two axle commercial
trucks
 Right: axles of trailers
Distribution of
estimated axle
spacings. There
are three clusters
in the data
separated by
empty bins. The
dotted lines
represent the
means of there
clusters
Outline
 Motivation
 Introduction
 Description
 Communication Protocol Design
 Experiment Setup
 Performance
 Conclusion & Future Work
Conclusion
 A novel algorithm that estimates the axle count and spacing





from pavement acceleration was designed and tested on the
collected data
ADET is simple enough to implement a sensor node with
limited processing power
Majorities of the existing technologies are wired solutions
Both the sensors and the AP are powered by batteries and
consume much less power than other technologies
The installation procedure and sensors themselves are
much cheaper
There is minimal maintenance compared to other
technologies
Future Work
 Find an optimal arrangement of sensors in order to
minimize the number of sensors deployed
 Reduce the amount of data transmitted
 Reduce the sensor power consumption