P ARK N ET : D RIVE - BY S ENSING OF R OAD -S IDE P ARKING S TATISTICS Suhas Mathur, Tong Jin, Nikhil Kasturirangan, Janani Chandrashekharan, Wenzhi Xue,

PARKNET: DRIVE-BY SENSING OF ROAD-SIDE PARKING STATISTICS

Suhas Mathur, Tong Jin, Nikhil Kasturirangan, Janani Chandrashekharan, Wenzhi Xue, Marco Gruteser, Wade Trappe

MobiSys ‘10

- Sowhat 2010.7.21

OUTLINES Motivation Objective Related Works ParkNet

Overview Hardware Deployment Detection Environmental positioning correction Viability

Discussion Conclusion

OUTLINES Motivation Objective Related Works ParkNet

Overview Hardware Deployment Detection Environmental positioning correction Viability

Discussion Conclusion

MOTIVATION Societal costs due to traffic congestion

Parking

Lack of information

Value of real-time information Government – adjusting prices Driver – suggested parking spaces, better

decision

OUTLINES Motivation Objective Related Works ParkNet

Overview Hardware Deployment Detection Environmental positioning correction Viability

Discussion Conclusion

OBJECTIVE

Providing parking information inslotted / unslotted areas

Parking information - Space Count Occupancy Map

OUTLINES Motivation Objective Related Works ParkNet

Overview Hardware Deployment Detection Environmental positioning correction Viability

Discussion Conclusion

RELATED WORKS SF-park project - Stationary sensor network

Sensor node installed in the center of each parking spot

Repeaters and forwarding nodes for connectivity Centralized parking monitoring system

Drawbacks of SF-park project Large installation and maintenance cost Limited to slotted parking spots Wireless relay nodes needed

OUTLINES Motivation Objective Related Works ParkNet

Overview Hardware Deployment Detection Environmental positioning correction Viability

Discussion Conclusion

PARKNET - OVERVIEW

Ultrasonic rangefinder GPS receiver

OUTLINES Motivation Objective Related Works ParkNet

Overview Hardware Deployment Detection Environmental positioning correction Viability

Discussion Conclusion

PARKNET - HARDWARE Ultrasonic sensors

Why ultrasonic? Low cost – compared to laser rangefinder Nighttime operation – compared to camera Increasing availability to support parking functions

Setting Range >= half the width of roads, 12~255

inches(0.3~6.5m) Sampling rate – several samples over the length of a car,

50ms

PARKNET - HARDWARE On-board PC

1GHz CPU 512MB RAM 20GB hard disk Atheros 802.11 a/b/g mini PCI card 6 USB 2.0 ports

GPS receiver Garmin 18-5Hz GPS 12 channel receiver 5 GPS readings / sec WAAS correction of errors less than 3 meters

Parking estimation server

OUTLINES Motivation Objective Related Works ParkNet

Overview Hardware Deployment Detection Environmental positioning correction Viability

Discussion Conclusion

PARKNET - DEPLOYMENT 3 vehicles 2 month 3 road-side area

(57 marked slotted parking spots/ 734m / 616m)

Total of more than ~500 miles of data Camera for ground truth GPS trip-boxes

Trigger to collect data Guard distance/time

OUTLINES Motivation Objective Related Works ParkNet

Overview Hardware Deployment Detection Environmental positioning correction Viability

Discussion Conclusion

PARKNET – DETECTIONSENSOR READING

PARKNET – DETECTIONFLOW & PRE-PROCESSING

Pre-processing Removing dips that have

too few readings < 6 sensor readings Assuming

max speed = 37 mph(58.2km)a car length = 5 m

PARKNET – DETECTIONFILTERING Comparing the width and depth of each dip

against thresholds

Threshold = the values make the minimum overall error rate

Overall error rate = false positive rate + miss detection rate

Depth >= 89.7 inches (2.27meters)Width >= 2.52 meters

PARKNET – DETECTIONWIDTH CLASSIFICATION Slotted

Dips of a width > 2*threshold 2 cars

UnslottedEstimating the spatial distance between

dips that have been classified as parked cars

Comparing the distance to the length of a standard parking space (6 meters)

PARKNET – DETECTIONMETRICS & EVALUATION Missed detection rate pm

False positive rate pf

Slotted – n’/ nUnslotted – d’/ d

OUTLINES Motivation Objective Related Works ParkNet

Overview Hardware Deployment Detection Environmental positioning correction Viability

Discussion Conclusion

PARKNET - ENVIRONMENTAL POSITIONING CORRECTION

GPS ERROR CORRELATION

PARKNET - ENVIRONMENTAL POSITIONING CORRECTION

ENVIRONMENTAL FINGERPRINTING

Comparing reported location of the pattern produced by fixed objects Priori known location of this pattern

Error vectorei(x,y) = ti(x,y) – li(x,y)

ti(x,y) : the true location of object i

li(x,y) : the location stamp of object i

Adding error vector to the location estimates of all detected cars within 100m of this object

PARKNET - ENVIRONMENTAL POSITIONING CORRECTION

EVALUATION Using Hungarian algorithm to match cars to

parking spots

OUTLINES Motivation Objective Related Works ParkNet

Overview Hardware Deployment Detection Environmental positioning correction Viability

Discussion Conclusion

Areas with greater amount of street parking utilization= areas with greater presence of taxis

Cost

Why ParkNet better? Nonguaranteed, random sampling

PARKNET - VIABILITY

ParkNet SF-park

Cost 400 / vehicle 250-800 / spot

Number 300 taxis 6000 spots

Total 120,000 1,500,000

OUTLINES Motivation Objective Related Works ParkNet

Overview Hardware Deployment Detection Environmental positioning correction Viability

Discussion Conclusion

DISCUSSION Multilane roads

Speed limitations

Obtaining parking spot maps

OUTLINES Motivation Objective Related Works ParkNet

Overview Hardware Deployment Detection Environmental positioning correction Viability

Discussion Conclusion

CONCLUSION ParkNet providing road-side parking info

Space countOccupancy map

300 taxis updating every ~25min for 80%

Some technical problems while implementing to real life

THANKS FOR LISTENING ~