2010 USDOT FHWA Travel Model Improvement Program (TMIP) National Webinar on Data Transferability and Data Mining - TRAC-IT: Travel Behavior Data Mining using GPS-enabled Mobile Phones
Download as PPTX, PDF1 like658 views
The document discusses 'Trac-It,' a mobile software designed for GPS-enabled phones to collect travel behavior data, offering features such as recording travel paths and providing real-time traffic alerts. It highlights the importance of battery life in GPS tracking, introduces data pre-processing and post-processing algorithms for efficient data collection, and describes methods for automated detection of transportation modes using neural networks. Additionally, it explores potential applications in real-time path prediction and integration with location-based services.
![© USF 2010
Patents Pending
Opportunities
• Proliferation of cell phones
– 61% of the world‟s population (4.1 billion) and 89% of
U.S. (276.6 million) are mobile subscribers (Jun. 09) [1][2]
– 23% of U.S. Households are Wireless–Only (Dec. 09) [3]
– E-911 mandate for locating cell phones
• Proliferation of cell phone “apps”
– While data is being collected from participant via phone,
location-aware mobile apps can provide services to user
(e.g. personalized traffic reports)
• Incentive for extended survey participation
• Longer survey period s with smaller samples for study
[1] International Telecommunications Union, “Measuring the Information Society - The ICT Development Index,” International
Telecommunications Union, 2009. [PDF]. Available: http://www.itu.int/ITU-D/ict/publications/idi/2009/material/IDI2009_w5.pdf. [
[2] http://www.ctia.org/media/industry_info/index.cfm/AID/10323
[3] http://news.yahoo.com/s/ap/20091216/ap_on_hi_te/us_cell_phones_only](https://image.slidesharecdn.com/barbeau-trac-it-dataminingnlg-130707180747-phpapp01/85/2010-USDOT-FHWA-Travel-Model-Improvement-Program-TMIP-National-Webinar-on-Data-Transferability-and-Data-Mining-TRAC-IT-Travel-Behavior-Data-Mining-using-GPS-enabled-Mobile-Phones-2-320.jpg)
2010 USDOT FHWA Travel Model Improvement Program (TMIP) National Webinar on Data Transferability and Data Mining - TRAC-IT: Travel Behavior Data Mining using GPS-enabled Mobile Phones
- 1. © USF 2010 Patents Pending TRAC-IT: Travel Behavior Data Mining Using GPS-Enabled Mobile Phones Sean Barbeau Research Associate Center for Urban Transportation Research University of South Florida National Centerfor Transit Research
- 2. © USF 2010 Patents Pending Opportunities • Proliferation of cell phones – 61% of the world‟s population (4.1 billion) and 89% of U.S. (276.6 million) are mobile subscribers (Jun. 09) [1][2] – 23% of U.S. Households are Wireless–Only (Dec. 09) [3] – E-911 mandate for locating cell phones • Proliferation of cell phone “apps” – While data is being collected from participant via phone, location-aware mobile apps can provide services to user (e.g. personalized traffic reports) • Incentive for extended survey participation • Longer survey period s with smaller samples for study [1] International Telecommunications Union, “Measuring the Information Society - The ICT Development Index,” International Telecommunications Union, 2009. [PDF]. Available: http://www.itu.int/ITU-D/ict/publications/idi/2009/material/IDI2009_w5.pdf. [ [2] http://www.ctia.org/media/industry_info/index.cfm/AID/10323 [3] http://news.yahoo.com/s/ap/20091216/ap_on_hi_te/us_cell_phones_only
- 3. © USF 2010 Patents Pending TRAC-IT • Mobile software for GPS-enabled phones – Like an iPhone App – It‟s OPT-IN • Features: – Runs on low to high tier phones – Records a person‟s travel behavior (an electronic activity diary) – Collects O/D and route information via GPS for all modes – Increases quality and quantity of collected information – Provides “hyper-personalized” real-time travel information services (e.g., traffic alerts)
- 4. © USF 2010 Patents Pending TRAC-IT • Two modes for TRAC-IT: – PASSIVE • No interactions with user, runs in background • Records GPS path, provides real-time services – ACTIVE • Adds questions at the end of their trips: – Name for location – Mode of Transportation – Purpose of Trip – Occupancy of Vehicle + TRAC-ITTRAC-IT <- Back Select (1) Work Related (2) Shopping (3) Pickup Someone (4) Go Home etc. ... Purpose of Trip:
- 5. © USF 2010 Patents Pending Assisted GPS data from TRAC-IT Asstd. GPS data from TRAC-IT
- 6. © USF 2010 Patents Pending GPS Data Pre-Processing • Battery life is key concern for mobile apps • If the user‟s phone dies, they will not use the app • Problems with tracking: – GPS consumes significant energy for each fix – Wireless communication drains battery fast • Solution: – Create data pre-processing algorithms that run on the cell phone before data is sent to server 6
- 7. © USF 2010 Patents Pending Impact of GPS on Battery 0 5 10 15 20 25 30 35 40 45 4 8 15 30 60 150 300 BatteryLife(hours) Interval Between GPS Fixes (seconds) Sanyo Pro 200 Sprint CDMA EV-DO Rev. A network 7
- 8. © USF 2010 Patents Pending Solution: GPS Auto-Sleep Sanyo Pro 200 Sprint CDMA EV-DO Rev. A network 0 20 40 60 80 100 120 140 160 180 200 1 21 41 61 81 101 121 141 161 181 201 221 241 261 281 301 321 341 361 381 401 421 441 461 481 501 521 541 561 581 601 621 641 661 681 701 721 741 761 781 801 821 841 861 881 901 921 941 961 981 1001 1021 TimeBetweenSequentialGPSFixes(s) “Asleep” “Awake” 8
- 9. © USF 2010 Patents Pending Impact of Wireless on Battery 9 Asstd. GPS data from TRAC-IT Sanyo 7050 Sprint CDMA 1xRTT Network UDP
- 10. © USF 2010 Patents Pending Asstd. GPS data from TRAC-IT Solution: Critical Point Algorithm 10
- 11. © USF 2010 Patents Pending GPS Data Post-Processing • Once the GPS data reaches the server, it is stored as records in a database – (X, Y) coordinates • In order to derive information from GPS, spatial data mining is necessary • Automation is key for large datasets! • Algorithms based on spatial operations can use spatial databases (e.g., PostGIS) 11
- 12. © USF 2010 Patents Pending Asstd. GPS data from TRAC-IT Hierarchical Clustering can find Points-of-Interest (POIs) 12 Points are clustered based on proximity to form a POI The remaining unclustered points naturally form discrete trips with Points of Interests (i.e., clusters) as starting and ending locations POI 1 POI 2POI 3 Trip 1 Trip 2
- 13. © USF 2010 Patents Pending Asstd. GPS data from TRAC-IT Ex. Car Trip BEFORE – Raw GPS data points AFTER - Points-of-interest identified as spatial regions Generates: –Trip start/end times & locations –Dwell times at POIs 13
- 14. © USF 2010 Patents Pending Ex. Walking trip 14
- 15. © USF 2010 Patents Pending Merging User POIs • Multiple visits to the same “location” should be registered with same POI • Needed to count visitation frequency • Algorithm uses POI overlap/bounding boxes to merge similar POIs • Can be per user, or aggregate 15
- 16. © USF 2010 Patents Pending Deriving Trip Characteristics from GPS data • Passive tracking places least burden on participant • However, surveyors often need additional data beyond GPS: – Mode of Transportation – Purpose – Occupancy • Can we automatically determine these characteristics just from GPS? 16
- 17. © USF 2010 Patents Pending Automated Mode Detection Bus Trip Car Trip Walking Trip Can artificial neural networks identify MODE from GPS data alone? 17
- 18. © USF 2010 Patents Pending Artificial Neural Networks • Two Step Process: – Training with known example data – Testing with new, unseen data Hidden Layers Outputs (Walk, Transit, or Car) Inputs (Speed, Acceleration, Estimated Accuracy, etc.) Weights Weights • AI tool for data-driven machine learning
- 19. © USF 2010 Patents Pending Sample Input Data Latitude Longitude Speed (m/s) Heading (0-359) Date & Time Est. Accuracy (m) Location Method 27.94330215 -82.33336639 13 286.52 2008-05-22 08:29:14.837 10.45 A-GPS 27.94348907 -82.33384704 7.75 292.34 2008-05-22 08:29:17.293 21.03 A-GPS 27.94371986 -82.33440399 5 298.57 2008-05-22 08:29:23.301 50.49 A-GPS 28.05500030 -82.40055847 - - 2008-05-22 08:29:26.529 - Cell-ID
- 20. © USF 2010 Patents Pending Input Data Attributes A-GPS vs. Cell-ID Cell-ID A-GPS Estimated Accuracy Uncertainty Calculated Location Potential True Location Latitude Longitude Accuracy Uncertainty (Radius in Meters) *Probability of approximately 68% 20
- 21. © USF 2010 Patents Pending Two Types of Datasets Studied All GPS Points Critical Points Only 21
- 22. © USF 2010 Patents Pending Choosing Data Input Attributes • User must choose data input attributes for neural network • Goal is to find attributes that will easily identify modes of transportation • Need to distinguish between similar modes – Especially Car vs. Bus 1222
- 23. © USF 2010 Patents Pending Choosing Data Input Attributes Good for Car vs. Bus Average Distance Between Critical Points 0 0.02 0.04 0.06 0.08 0.1 0.12 1 2 3 4 5 6 7 Trip Sample AverageDistancebetweenCriticalPoints (miles) Car Walk Bus 1623
- 24. © USF 2010 Patents Pending Final Inputs • Inputs chosen for All GPS Points dataset: – Avg. Speed – Max. Speed – Avg. Accuracy Uncertainty – Percent Cell-ID Fixes – Standard Deviation of Distance Between Stop Locations – Average Dwell Time 17
- 25. © USF 2010 Patents Pending Final Inputs • Inputs chosen for Critical Points Only dataset: – Avg. Speed – Max. Speed – Avg. Acceleration – Max. Acceleration – (# of Critical Points / Total distance of the trip) – (# of Critical Points / Total time of the trip) – Total Distance – Average Distance between critical points 1825
- 26. © USF 2010 Patents Pending Experiment • 114 trips recorded in Tampa, Fl – 38 car – 38 bus – 38 walk • Devices = Motorola i870 and i580 phones – Sprint-Nextel iDEN network • Software = TRAC-IT mobile app. – Java Micro Edition w/ JSR179 Location API – Queries position every 4 seconds 2026
- 27. © USF 2010 Patents Pending Experiment • Neural Network Software = Weka – Used Java API for Multi-Layer Perceptron • 10-fold cross validation used – Full data set randomly partitioned into 10 sets – 9 sets used for training, 1 set for testing – Repeated 10 times while alternating testing set – Reported accuracy is mean value of 10 tests 2127
- 28. © USF 2010 Patents Pending Results • Numerous neural network settings were tested • Best results: Type of Input Accuracy All GPS Points 88.6% Only Critical Points 91.23% -Using .1 Learning Rate and 300 training epochs Good for mobile phone battery! 2228
- 29. © USF 2010 Patents Pending Results Mode of Transportation Average Accuracy Per Mode Car 92.11% Bus 81.58% Walk 100.0% • Breakdown of 91.23% accuracy for Critical Point Only dataset Similar traits 29
- 30. © USF 2010 Patents Pending • Use GIS Land-Use and Zoning maps to determine location type – Single-Family Home – Restaurant – Etc. • Derive purpose from location type Automated Purpose Detection
- 31. © USF 2010 Patents Pending • Used Hillsborough County Department of Revenue (DOR_CD) and Zone_Prime fields to programmatically identify land use Proof-of-concept Purpose Detection
- 32. © USF 2010 Patents Pending USE CODE PROPERTY TYPE Residential – 0000 Vacant Residential – 0100 Single Family – 0200 Mobile Homes Commercial – 1300 Department Stores – 1400 Supermarkets – 1600 Community Shopping Centers – 1700 Office buildings, non-professional service buildings, one story – 2000 Airports (private or commercial), bus terminals, marine, etc. – 2100 Restaurants, cafeterias – 2200 Drive-in Restaurants 2300 Financial institutions (banks, savings and loan companies, etc.) – 2400 Insurance company offices 2500 Repair service shops (excluding automotive) Institutional 7100 Churches – 7200 Private schools and colleges – 7300 Privately owned hospitals – 7400 Homes for the aged – 7500 Orphanages, other non-profit or charitable services Sample DOR Codes …. …. ….
- 33. © USF 2010 Patents Pending Automated Purpose Detection • Possible, but many challenges: – Multi-use areas: • Baseball field at a school – Alternate uses: • Work at a restaurant – Coding issues: • Red Lobster designated as “Federal” instead of restaurant • Likely useful for prompted recall – ~65% accurate in proof-of-concept
- 34. © USF 2010 Patents Pending TRAC-IT User ALERT! You are headed towards an accident at S. Howard and Crosstown Expy • “Hyper-personalized” real-time traffic alerts Provide Value to Participant 34
- 35. © USF 2010 Patents Pending What is Path Prediction? • Real-time spatial data mining • Predicts a user‟s real-time path using: – Real-time location – Historical travel behavior • Based mainly on spatial data operations • Once path is predicted, algorithm can find alerts along a traveler‟s predicted path – Ex. Traffic accidents, advertising • Reduces irrelevant alerts sent to users 35
- 36. © USF 2010 Patents Pending How It Works • Two steps: – Part A - Build user history over time from traveled paths – Part B – Predict immediate travel behavior based on real- time and historical travel behavior 36
- 37. © USF 2010 Patents Pending How It Works (Part A) As user travels over time, recorded GPS data is translated into paths (polygons) in database GPS data points 37
- 38. © USF 2010 Patents Pending How It Works (Part A) As user travels over time, recorded GPS data is translated into paths (polygons) in database Converted to polyline… 38
- 39. © USF 2010 Patents Pending As user travels over time, recorded GPS data is translated into paths (polygons) in database How It Works (Part A) Converted to polygon. 39
- 40. © USF 2010 Patents Pending In real-time, phone sends GPS fixes to TRAC-IT server… How It Works (Part B) 40
- 41. © USF 2010 Patents Pending …Server runs Path Prediction and checks path history… How It Works (Part B) Path Prediction 41
- 42. © USF 2010 Patents Pending How It Works (Part B) 42 …An algorithm using a series of spatial operations identifies the most likely paths…
- 43. © USF 2010 Patents Pending Server checks for incidents intersecting predicted paths from real-time data source, and alerts phone How It Works (Part B) 43
- 44. © USF 2010 Patents Pending 44
- 45. © USF 2010 Patents Pending 45
- 46. © USF 2010 Patents Pending 46
- 47. © USF 2010 Patents Pending Current Path Prediction Work • Integrate Bayesian predictions based on POI visitation frequency with spatial predictions • System integration with real-time travel information data sources in Florida 47
- 48. © USF 2010 Patents Pending Acknowledgements • Center for Urban Transportation Research (CUTR) – Phil Winters, Nevine Georggi • USF Computer Science Department – Miguel Labrador, Rafael Perez • National Center for Transit Research • Florida Department of Transportation • US Department of Transportation • National Science Foundation • Sprint-Nextel Application Developer Program 48
- 49. © USF 2010 Patents Pending Questions? Research Associate Center for Urban Transportation Research University of South Florida (813) 974-7208 USF Location-Aware Information Systems Lab: http://www.locationaware.usf.edu/ Sean J. Barbeau, M.S. 49
- 50. © USF 2010 Patents Pending For Additional Reading… • Paola A. Gonzalez, Jeremy S. Weinstein, Sean J. Barbeau, Miguel A. Labrador, Philip L. Winters, Nevine L. Georggi, Rafael A. Perez. “Automating Mode Detection for Travel Behavior Analysis by Using GPS- enabled Mobile Phones and Neural Networks,” Institution of Engineering and Technology Intelligent Transportation Systems Journal. doi: 10.1049/iet-its.2009.0029 (to appear 2010). • Sean J. Barbeau, Nevine L. Georggi, Philip L. Winters. “TRAC-IT: Travel Behavior Data Collection using GPS-enabled Mobile Phones,” Human Factors 135 F – Quantifying Driving-Risk Exposure Committee Meeting at National Academy of Sciences’ Transportation Research Board 89th Annual Meeting. Washington, D.C., January 9th, 2010. • Sean J. Barbeau, Miguel A. Labrador, Nevine L. Georggi, Philip L. Winters, Rafael A. Perez. “TRAC-IT: A Software Architecture Supporting Simultaneous Travel Behavior Data Collection and Real-Time Location- Based Services for GPS-Enabled Mobile Phones,” Proceedings of the National Academy of Sciences’ Transportation Research Board 88th Annual Meeting, Paper #09-3175. January, 2009. • Narin Persad-Maharaj, Sean J. Barbeau, Miguel A. Labrador, Philip L. Winters, Rafael Perez, Nevine Labib Georggi. “Real-time Travel Path Prediction using GPS-enabled Mobile Phones,” 15th World Congress on Intelligent Transportation Systems, New York, New York, November 16-20, 2008. • Sean J. Barbeau, Miguel A. Labrador, Philip L. Winters, Rafael Perez, Nevine Labib Georggi. “Trac-It - A „Smart‟ User Interface For A Real-Time, Location-Aware, Multimodal Transportation Survey,” 15th World Congress on Intelligent Transportation Systems, New York, New York, November 16-20, 2008. • Paola A. Gonzalez, Jeremy S. Weinstein, Sean J. Barbeau, Miguel A. Labrador, Philip L. Winters, Nevine Labib Georggi, Rafael Perez. “Automating Mode Detection Using Neural Networks and Assisted GPS Data Collected Using GPS-Enabled Mobile Phones, 15th World Congress on Intelligent Transportation Systems, New York, New York, November 16-20, 2008. • Sean J. Barbeau, Miguel A. Labrador, Alfredo Perez, Philip Winters, Nevine Georggi, David Aguilar, Rafael Perez. “Dynamic Management of Real-Time Location Data on GPS-enabled Mobile Phones,” Presented at UBICOMM 2008 – The Second International Conference on Mobile Ubiquitous Computing, Systems, Services, and Technologies, Valencia, Spain, September 29 – October 4, 2008. © 2008 IEEE. http://www.locationaware.usf.edu/publications.htm
Editor's Notes
- #18: What if GPS data identified MODE?Use neural networks
- #28: Number of hidden nodes = (Number of attributes + Number of possible classifications) / 2
- #29: Number of hidden nodes = (Number of attributes + Number of possible classifications) / 2
- #35: Personalized Travel Information Provides User with Direct Participation Incentive