Signal guru

SignalGuru: Leveraging Mobile
Phones for Collaborative Traffic
Signal Schedule Advisory
Emmanouil Koukoumidis , Li- Shiuan
Peh, Margaret Martonosi
Princeton University
MIT

BOURSINOS CHRISTOS/DISTRIBUTED SYSTEMS/ Msc CS FALL 2011/V. KALOGERAKH

1

Traffic conditions
Traffic signals despite that they are a safety control
1. Enforce a stop and go movement
2. Increases fuel consumption
3. Reduces traffic flow
4. Traffic jams
For all problems that we saw before we have some solutions
1. Countdown timer at vehicular traffic signals ( Komotini,
Alexandroupoli)
2. Countdown timers for pedestrian traffic signals(USA)
3. GLOSA
results from CAMBRIDGE and SINGAPORE.

2

We can alleviate this problem with computational devices
3. GLOSA (Green Light Optimal Speed Advisory)
•

•
•
•

detect current traffic signals with their camera
collaboratively communicate
Learn traffic signal results patterns
Predict their future schedule

AUDI recently prototyped a small scale DSRC-based GLOSA
system for 25 traffic signals in Ingolstadt
It is expensive to equip all the cars with such technology.
BOURSINOS CHRISTOS/DISTRIBUTED
SYSTEMS/ Msc CS FALL 2011/V.
KALOGERAKH

3

Where we put our system in our car.

4

GLOSA faces several challenges:
1. Lack of Loop detector information ( traffic signals based on
information from loop detectors embedded under every lane
on roads close to the stop line governed by traffic signals)
2. Commodity cameras ( the quality of smart phones cameras)
3. Limited processing power ( takes significant computational
resources)
4. Uncontrolled environment composition and false detection
(no control over the composition of the content captured by
their video cameras)
5. Variable ambient light conditions
6. Need for collaboration ( not be able to see a far-away traffic
signal, or may not be within view of the traffic signal for a
long enough stretch of time)

5

GLOSA
The goal of GLOSA is to advise drivers on the optimal Speed they
should maintain so that the signal is green when they arrive at
the next intersection.
What offers
1. Decreased fuel consumption
2. Smoothed and increased traffic flow
3. Decrease environment impact
Need four pieces
1. The residual amount of time till the traffic signal changes
2. The intersection location(map)
3. Vehicles correct location( gps )
4. The queue length of the traffic ahead.
May improve an individual vehicles travel time.

6

Other Possible SignlaGuru-Enabled applications
1. Traffic Signal-Adaptive Navigation(TSAN)
2. Red Light Duration Advisory(RLDA)
3. Imminent Red Light Advisory(ILVA)
4. Red Light Violation Advisory (RLVA)
In the next slide we see the deference between TSAN
and GLOSA.

KALOGERAKH

7

Compare GLOSA-TSAN
• The deference
is that in
GLOSA we
have the
optimal speed
and in TSAN
we have a
suggestion
detour.
• Architecture of
GLOSA

KALOGERAKH

8

SIGNALGURU ARCHITECTURE
Detection Module/ detection algorithm
• Which color
qualify
• Best candidate
to be a traffic
signal
• BCC percentage
of the pixels fall
into the correct
color range
• How many pixel
are dark enough
to qualify as
traffic signal

9

How the device do the detection. COLOR FILTER


10

IMU-based Detection Window


11

With what angles do the detection

• Angle θ=φ/(2-χ) angle χ =ψ-ω angle ψ=arctan(hs-hc)/d
KALOGERAKH

12

Variable Ambient Light Conditions
Our program will not perform good in some cases
1. Time of day
2. Weather conditions
To solve this problem we change the sensitivity of the mobile
camera to more sensitive.
We have some buttons to do this job in our program


13

Transition filtering module
filters
We filter RG transition using a two-stage filter:
1. Low Pass Filter(LPF) in 1st stage
2. Collocation filter in 2nd stage
LPF 1st stage
In 88% of the cases, false positive detections occur over a single
frame and do not spread over multiple consecutive frame:
1. R…RGR…R
When a car is waiting at the red light it correctly detects, then at
a specific instance it misdetects a passing object for a green
traffic light.

14

LPF 1st stage
2. G…GRG…G
When the vehicle misdetects an arbitrary object for a red light
between detections of the actual green light.
3. NS…NSRGNS...NS
When the view of the car is obstructed and there is no traffic
signal in sight. However in some point it misdetects an
arbitrary object for a R and G light.
The LPF classifies only this transmission
RRGG
Colocation filter 2nd stage
Checks whether the green bulb that was just detected is close to
the red bulb detected in the previous frame.

15

Collaboration module
A node(cell phone) needs information about a traffic signal well
before the signal comes into the node’s camera field.

The collaboration module allows participating SignalGuru nodes
to opportunistically exchange their traffic signal information
by periodically broadcasting UDP packets in 802.11 ad-hoc.
So they predict the schedule by using a database of the traffic
signal settings for a period of time.


16

PREDICTION MODULE
We have two main categories of traffic signal:
1. Pre-timed traffic signals
2. Traffic-adaptive traffic signals
Because their operation is very different , SignalGuru uses
different prediction schemes for each category


17

Pre-timed traffic signals
SignalGuru’s prediction module maintains a database of the
traffic signal settings.
This means that SignalGuru knows how long each phase lasts.
The Challenge is to synchronize the SygnalGuru clock with the
time of phase transition of a traffic signal.


18

Pre-timed traffic signals
Light switch to red (phase A) switch to Green (phase B)
Phase A will follow after phase B.
If we have a false detection the synchronize needs to be
reestablished.


19

Traffic-adaptive traffic signals
SignalGuru predicts future transitions by detecting past
transitions and predicting the length of the current of next
phases.
The key different from the prediction of pre-timed traffic signals
lies in the prediction of the phase length, as opposed to
looking it up from a database.
SignalGuru predicts the length of a phase by measuring and
collaboratively collecting the prior traffic signal transition
history and feeding it to a Support Vector Regression(SVR)
prediction model.
We evaluate the prediction performance of different Prediction
Schemes(PS) by training the SVR with different sets of
features.

20

Traffic-adaptive traffic signals
One-week-history long of data is enough to train the SVR model.
Furthermore the SVR model does not need to get continuously
re-trained. Re-training the model every 4 to 8 months is
frequent enough in order to keep the prediction errors
small.


21

METHODOLOGY
CAMBRIDGE DEPLOYMENT
• 5 cars with i-phone and
the drivers follow the
rout.
•

One more device P2
pedestrian.

•

P2 was the ad-hoc data
relay node.


22

METHODOLOGY
SINGAPORE DEPLOYMENT
•
•
•
•
•

•

2 cars following two routs.
8 i-phones in taxis
5 mobile in rout A
3 mobile in rout B
One more device P2
pedestrian.
P2 was the ad-hoc data
relay node. Also recording
the ground truth when the
traffic signals status
transitioned.


23

SIGNALGURU EVALUATION
TRAFFIC SIGNAL DETECTION
We evaluate the performance of
two deployments.

•
•

5959 frames Cambridge
1352 frames Singapore

•
•

False 7,8% Cambridge
False 12,4% Singapore

•
•

Correctly 92,2% Cambridge
Correctly 87,6% Singapore


24

We evaluate the benefits that
IMU-based detection
window offers.
•

•

•

The IMU reduces significantly
the number of red false
positives.
Often confuses vehicles for a
light.
Reduces the average
processing time by 41%


25

Transition Filtering
The performance of the
transition Filtering.
The False Positive(FP) in
Cambridge is smaller than
in Singapore
1. Rate of FP traffic signal
detection is smaller in
Cambridge
2. The average waiting
time at red traffic signals
is only 19,7s in C. vs
47,6s in S.

26

Schedule Prediction
Cambridge deployment
• Pre-timed with average
error 0,66s
• Video frame every T=2sec
• Εmax=T/2
• EXP=T/4=o,5 sec
• Can effectively support
the accuracy
requirements of all
applications.

27

Schedule Prediction
Singapore deployment
• Average error 2,45 sec


28

GLOSA Fuel Efficiency
• From OBD-LINK

• FROM P1P2
• Reducing fuel
consumption on average
by 20,3%
• Improves the vehicle’s
mileage on average by
24,5%


29

Related works
• In Lee “et al” propose an application that lets police track
the movement of suspicious vehicles based on information
sensed by camera-equipped vehicles.
• Other works proposed to equip vehicles with specialized
cameras and detect traffic signals with the ultimate goal of
enabling autonomous driving , assisting the driver, detecting
the location of intersection and overlaying navigation
information.
• Our System GLOSA putting the IMU and also it has safe
results. It use DLT certificates or a TPM in order to ensure
trust in the exchange of traffic signal data.


30

Conclusions
• In order to predict traffic signals future schedule and
support a set of novel applications is a fully distributed and
grassroots approach.
• Our proposed schemes improve traffic detection filter noisy
traffic signal data and predict traffic signal schedule.
• SignalGuru can effectively predict the schedule for not only
pre-timed but also state of the art traffic-adaptive traffic
signals.
• Fuel efficiency

We hope for a motivation from you.

31

The End


32

Signal guru

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Similar to Signal guru

Similar to Signal guru (20)

Recently uploaded

Recently uploaded (20)

Signal guru