SLIDE 1
6.808: Mobile and Sensor Computing
Lecture 10: The Pothole Patrol
Hari Balakrishnan hari@mit.edu
Slides from Jakob Eriksson
6.808: Mobile and Sensor Computing Lecture 10: The Pothole Patrol - - PowerPoint PPT Presentation
6.808: Mobile and Sensor Computing Lecture 10: The Pothole Patrol - - PowerPoint PPT Presentation
6.808: Mobile and Sensor Computing Lecture 10: The Pothole Patrol Hari Balakrishnan hari@mit.edu Slides from Jakob Eriksson road decay unavoidable, hard to predict current monitoring methods costly/ineffective 3 the Pothole Patrol
SLIDE 2
SLIDE 3
- road decay unavoidable, hard to predict
- current monitoring methods costly/ineffective
3
SLIDE 4
the Pothole Patrol
4
' P
- t
h
- l
e d e t e c t e d a t . . . . '
Open WiFi Access Point
GPS localization P2 Central Server
- pportunistic accelerometer sensing
GPS localization
- pportunistic data upload
aggregation and reporting
3-axis accelerometer (380 Hz)
SLIDE 5
Acceleration vector
5
ay: road direction ax: on road plane, perpendicular to road az: perpendicular to road plane Road
SLIDE 6
experimental platform
- 7 Boston/Cambridge taxis
- small computer in glove box
- 380 Hz 3-axis accelerometer
- 802.11a/b/g wireless interface
- GPS receiver on roof
- <time,location,heading,speed,ax,ay,az>
6
SLIDE 7
7
wide-area sensing
SLIDE 8
8
SLIDE 9
9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 10 100 1000 Fraction of road segments Number of repeat encounters
7 cars in 10 days 2492 unique km 9730 total km appear @ least once
SLIDE 10
P2 architecture
10
3-Axis accelero- meter
GPS Location Interpolator Pothole Record Sensors Clustering
x x x x x x xx
1.5 1.1 2.6
Pothole Detector Vehicle clients Central server
SLIDE 11
sensor placement
- highly accurate
- difficult mounting
- extreme exposure
try to stay inside vehicle Pros? Cons?
SLIDE 12
12
480 500 520 540 560 580 600 371 372 373 374 375 376 Acceleration (raw) Time into trace (sec) Attached to Windshield
- very clean signal
- ‘gold standard’
- difficult to mount
490 500 510 520 530 540 550 560 570 368 369 370 371 372 373 Acceleration (raw) Time into trace (sec) Attached to Dashboard
- good signal
- easy to mount
- ‘out of the way’
- 40
- 20
20 40 60 80 354 355 356 357 358 359 Acceleration (raw) Time into trace (sec) Attached to Embedded PC
- very poor signal
- no mounting necessary
WINDSHIELD?
DASHBOARD?
ATTACHED TO PC?
SLIDE 13
challenge: “pothole” v. “not pothole”
13
How do I identify pothole vs others?
SLIDE 14
pothole v. not pothole
14
SLIDE 15
SLIDE 16
16
high-pass speed
windows
- f all event
classes door slams, curbs
z-peak xz-ratio speed vs. z ratio
acceleration, braking, turns pothole detections minor anomalies expansion joints rail crossings speed bumps smaller highway anomalies
OUT IN
P2 detector
256-sample windows
Events usually of much shorter duration than 256 samples
SLIDE 17
17
high-pass speed
windows
- f all event
classes door slams, curbs
z-peak xz-ratio speed vs. z ratio
acceleration, braking, turns pothole detections minor anomalies expansion joints rail crossings speed bumps smaller highway anomalies
OUT IN
P2 detector
256-sample windows
Events usually of much shorter duration than 256 samples
Need to learn threshold parameters (will come back to it later)
SLIDE 18
18
SLIDE 19
x-z ratio “high enough”
19
SLIDE 20
hand-labeled training data
20
SLIDE 21
training the detector
21
Type Count Percentage Smooth road (SM) 64 23% Potholes (PH) 63 23% Manholes (MH) 59 21% Railroad Crossing (RC) 18 6% Crosswalk/Exp. Joint (CWEJ) 76 27%
- manually label training samples
SLIDE 22
loosely-labeled training
- needed to avoid over-training with
unrepresentative manually curated data
- under-samples “smooth” roads
22
SLIDE 23
training the detector
- pick an objective function
- ptimize over 3 threshold parameters
- z-peak
- xy-ratio
- speed vs. z-ratio
23
For each set of parameter score s(t) = corr − incorr2. detections reported when using
SLIDE 24
detector performance
24
After training on loosely labeled data E.g., 7.3% of detected “potholes” are railroad
Note: Actual false positive rate is not 7.6% Why?
Among the segmented reported as potholes by the algorithm
SLIDE 25
25
Road # potholes #win #det. rate Storrow Dr. few 1865 3 0.16% Memorial Dr. few 1781 2 0.12% Hwy I-93 few 2877 5 0.17% Binney St some 6887 25 0.63% Beacham St many 1643 231 14%
estimating false +ve rate
# of sample windows # of detections/ # windows upper bound on FPs
SLIDE 26
clustering
26
SLIDE 27
- 7 taxis over 10 days
- 9730 total km of road covered
- 2492 unique km of road covered
- 1.4 million sample windows
- 4131 severe detections in 2709
locations (after clustering)
27
experiments
SLIDE 28
28
SLIDE 29
29
potholes 39 sunk-in manholes 3 railways and exp. joints 4 undetermined 2
48 spot-checks
SLIDE 30
30
SLIDE 31
SLIDE 32
- automatic wide-area road
quality monitoring
- use of opportunistic mobility
- mobile sensing w/ delay-tolerant
communication
- machine learning classifier with labeled and
loosely-labeled data
- Data collection and curation is hard!
- low-cost approach to help solve
a costly problem