Transcript Kyun Queue - Network and Systems Laboratory

Kyun Queue:
A Sensor Network System to Monitor
Road Traffic Queues
2012.09.24 Junction
Rijurekha Sen, Abhinav Maurya, Bhaskaran Raman, Rupesh
Mehta, Ramakrishnan Kalyanaraman, Nagamanoj
Vankadhara, Swaroop Roy, Prashima Sharma
India Institute of Technology, Bombay
Abstract
• A SWN system for real-time traffic queue monitoring.
• Works in chaotic traffic
• Doesn’t interrupt traffic flow during its installation and
maintenance
• Incurs low cost
Introduction
• Road congestion/traffic queue
• Unpredictable travel times
• Fuel inefficiency
• Non-lane base and highly heterogeneous
Contribution
• Propose a new mechanism to sense road occupancy based on
variation in RF link characteristics, when line of sight between
a transmitter-receiver pair is obstructed
• Algorithms to classify traffic states into congested or freeflowing at time scales of 20 seconds with above 90% accuracy
• This network can correlate the traffic state classification
decisions of individual sensors, to detect multiple levels of
traffic congestion or traffic queue length on a given stretch of
road, in real time.
• Deployment of our system on a Mumbai road, after careful
consideration of issues like localization and interference, gives
correct estimates of traffic queue lengths, validated against 9
hours of image based ground truth.
Other solutions
Infrared sensors: overly sensitive to even small obstacles
Traffic sensing
25m
• 802.15.4 compliant Telosb motes
• Tx sends 25 pkt/sec, payload 100bytes at -25dBm power
Real-time classification of traffic states
• Window size = 5 minutes?
• Adaptive traffic light control would intuitively need faster inputs
• 1 minute cycle time
• Very low time windows give noisy predictions
• the inherent stochastic nature of wireless links which causes link
quality to be intermittently bad though the tx-rx are in perfect
LOS
• the instantaneous traffic condition between the tx-rx are contrary
to the actual traffic state
Labeled data-set for evaluation
• 16 hours data-set from
• 25m wide Adi Shankaracharya Marg (wide)
• 13676 seconds: free-flow
• 14992 seconds: congested
• 8m wide henceforth (narrow)
• 13486 seconds: free-flow
• 16678 seconds: congested
• Labeling
• Larger time-scale of 5 minutes to reduce manual overhead
• Traffic states did not toggle within 5 minutes
Classification Algorithm
• Machine learning
• Binary traffic state classification
• (a) congested traffic
• (b) free-flowing traffic
• Trade-off between
•
•
•
•
(a) accuracy of classification
(b) implementability on a low end embedded platform
(c) complexity of the classifier models
(d) overhead of model training
• Two possible solutions
• (a) FeatureClassifier (FC) algorithm
• (b) SignalClassifier (SC) algorithm
Which combination?
• FeatureClassifier (FC) algorithm
• Feature vector comprising 9 features (RSSI percentile)
10th
20th
30th
40th
50th
• SignalClassifier (SC) algorithm
• Voting with one RSSI
• SVM, K-means
60th
70th
80th
90th
Choice of features
• LQI (link quality index), RSSI, PRR
• Only use RSSI is better
• Non-intuitive observation : RSSI is much more strongly correlated
with line-of-sight than LQI or PRR
Choice of features
• There are several IEEE 802.15.4 compliant radios like XBEE,
which do not report LQI value.
RSSI shows good difference
between the two traffic
states even for small d’’.
Classification models
• Use FeatureClassifier algorithm with K-Means model
• Built over 20 seconds of RSSI percentiles
• Training model is built from the 8 hour dataset collected from
roads
Design of the Kyun system
• Architecture
• 3 pairs of transmitter (Ti) – receiver (Ri) perform sensing and
computation to know the traffic condition.
• A central controller unit (C) resides on the traffic light. C upon
receiving the road occupancy observation values from Tx-Rx pairs can
compute the optimal green light distribution
• C can communicate with a server (S). S upon receiving road
occupancy information can implement other applications like
coordinated signal control, bottleneck identification and congestion
mapping
Which RF to use?
• 802.11 a/b/g requires higher power
• Bluetooth shows an issue of poor range and links are hard to
establish for wide roads of 25m.
• Choose 802.15.4
Sensing and communication confilicts
• (1) sensing links: across the road from T and R
• 0.5 m from the ground level
• (2) communication links: along the road, from on R to R
• R is mounted on the road-side Lamp-post (30m)
• Fault: alternate lamp-posts to communicate => 60 m
• Experiment
•
•
•
•
T: transmit 25 pkt/sec
30m:10m:100m
6 am: 60m at 60% PRR
8 am: >30m at 40% PRR
Resolution
• Two 802.15.4 radios in R units
• XBEE radio for sensing (0.5m)
• CC2520 radio for communication (2m)
• Clear of pedestrians on the footpath
Software protocol
• Ensure that all R’s perform sensing and computation
simultaneously
• C centrally controls measurement cycle
• No need of any explicit time synchronization mechanism
•
C-RDY: receiver ready mode
for a control message
MACwaiting
protocol
• Simple CSMA-CA with 4 tries
C sends a control message
1. Write in micro-SD card and
Compute the traffic signal
schedule from the queue
length information in the
data message (every 30s)
2. Send to S by GPRS
Compute sensing decision
Decision results originated
by the last R
Hardware prototypes
$200
$250
Deployment based evaluation
• Classification accuracy
• 50 minutes
• 149/150 correct
• 99.67%
Deployment based evaluation
• Deployment
C
3
R1
R2
1
1
R3
R4
R5
1
0
0
• Nov. 17(Thu), 18(Fri), 19(Sat), 2011 6pm – 9pm
• Image based manual verification scheme
• Samsung Google Nexus phone
to capture image ever 30 seconds
• Can only cover T-R pair 1-3
• 1 person observes offline and estimates
the queue length
• 2nd person as consulter
Deployment based evaluation
• Only cover 0-3, error up to 3 units
Deployment based evaluation
• The queue buildup and clearing was very rapid on that day
Self localization
• Without pre-defined order and program R unit
• To use RSSI ranging
• It is not a monotonically decreasing function of distance
• Two-ray model equation: lower antenna heights seem to give less
oscillations
Self Localization
• Vertical polarization seems to give far less oscillations than
horizontal polarization
• Horizontal: {2, 3, 5, 6, 4}
Vertical: {2, 3, 6, 5, 4}
• Reordering occur independent of antenna orientation and polarization
• Vertical orientation gives worse RSSI values
• Fading effect of the metal body of the lamp-post with which the vertical
antenna is parallel
Self Localization
• Using different 802.15.4 channels, the average RSSI over the
channels would have less oscillations
• Useless
• None of the parameters
like antenna height, polarization
or channel, seem to be useful to
remove the RSSI oscillation
Self Localization
• Fig. 30: {node 2, node 5, node 4, node 6, node 3}
31
70
95
116
144
• Fig. 31: {node 2, node 5, node 6, node 4, node 3}
39
64
85
113
140
• The R recording the best RSSI is marked as the first node
Discussion and future work
• Interference issues
• 802.15.4 has 16 orthogonal channel
• Inter-network interference can still exists (Wi-Fi AP)
• Interference doesn’t affect RSSI of received packets
• Disable CRC
• Disable CCA
• Use only 802.15.4 channel 26
• Outside 802.11 spectrum
Discussion and future work
• Power consumption
• T units: 20 XBEE tx operations per second
• R units: in every measurement cycle spanning 30 seconds,
• receive at most 400 XBEE packets during sensing
• Perform one classification operation
• Receive and transmit at most 8 CC2520 messages
• C unit:
• performs 1 GPRS communication
• 2 SD card operations
• Receive and transmit at most eight CC2520 messages
Discussion and future work
• Classification schemes and model training
• For different kinds of roads
• Semi-supervized training – balance between overhead of manual
labeling in supervized learning and noise in unsupervized
methods
• Identify characteristics like road width and vehicle types, that
affect model parameters, and subsequently see if roads similar in
those characteristics can use the same model with sufficient
accuracy
• Extend -> muticlasses
•
•
•
•
Empty road
Fast traffic
Slow traffic
Standing traffic
Discussion and future work
• Other design choices
• Different topologies to reduce overhead
• Replace dual radio solution with single radio
• …..
Conclusion
• Design and implement a new sensing system to detect road
occupancy based on RF link quality degradation
• Design and implement a sensor network to distributedly
decide traffic queue length in real time
• Achieve upto 96% accuracy in queue length estimation