| City: |
DC/Maryland/Virginia, U.S.
|
| Organization: |
Noblis; Department of Civil and Environmental Engineering, Maryland Transportation Institute, University of Maryland
|
| Project Start Date: |
Unknown
|
| Project End Date: |
18 June, 2018
|
| Reference: |
Krause, Cory M., and Lei Zhang. “Short-Term Travel Behaviors Prediction with GPS, Land Use, and Point of Interest Data.” Transportation Research Part B, vol. 123, May 2019, pp. 349–361., Elsevier Ltd. doi:10.1016/j.trb.2018.06.012.
|
| Problem: |
Traffic congestion is causing significant inefficiency in daily commute and waste in fuel in the U.S. While tons of traveler information are collected through many federal projects that utilize these information and intelligent transportation systems to alleviate congestion, there had not been any tools developed to help manage route selection and travel demand in real life by predicting vehicles’ destinations at beginning of trips with estimated trip purposes.
|
| Technical Solution: |
Firstly, trip purposes were predicted. The first 15 trips taken by each participant were recorded and put into a pre-established rule system to generate the purpose for each of those trips. Using the trip purposes developed through the first 15 trips and the set of rules developed, the future trips will be assigned a purpose based on only start-of-trip information.
Then, the trip information at beginning of trips including trip time and day of week, origin, purpose, user and trip number are put into a Hierarchical Markov Model (HMM). The model searches among the previous trips in the Tier Array from defined Tier 1 attribute to tiers below until an estimation on the trip destination can be made based on the percentage of trips found in previous trips that ended in the same destination with same attribute values. The model adopted a tier structure of purpose, time and origin, time, origin, most common destination (from top to bottom). If no estimation could be made due to lack of information from previous trips or no higher percentage of occurrence for any destination comparing to other destinations, the model will move on to the next lower tier. Once a trip completes with an actual destination, the trip information including the actual destination will be added to the Tier Array for estimating the next trip’s destination.
|
| Datasets Used: |
- Dataset 1: GPS driving traces; University of Maryland in conjunction with U.S. Federal Highway Administration and Maryland Highway Administration; October 2011 to February 2012
- Dataset 2: point of interest (POI)/land use data; OpenStreetMap; unknown date
|
| Outcome: |
The models developed in predicting trip destinations prior to this model did not take into consideration of land use patterns; this model utilizes land use data to estimate trip purpose which significantly increased the accuracy in real-time destination prediction by 7% on average for all types of trip purposes, and by 15% for some specific types of trips. For work-purposed trips, the model has an accuracy of over 90% in predicting the destination. This result was measured after just a short period of data learning, it is believed that the model’s predictive accuracy will be improved by keep learning traveller behaviors.
|
| Issues that arose: |
- Issue 1:Due to the computational time of linking POI and GPS points, POI was only linked to GPS points within the DC/Maryland/Virginia regions. The long-distance trips included in this research do not have land use data allocated to them, which made it hard to predict trip purposes for long-distance trips. This may reduce the accuracy of the model, however, due to relatively small number of long-distance trips included in the dataset, the impact is believed to be minor.
- Issue 2:A larger training set for trip purpose learning would increase the accuracy in predictions in long run, however, will sacrifice the trips’ predicted accuracy included in the learning set. It is possible to update the trip purpose definition after each trip is taken, however, will require significant computing power.
|
| Status: |
Terminated, but the model can be applied to future work in en-route prediction of destination.
|
| Entered by: |
Date: October 30, 2020: Miranda Shuang Liu, smiranda.liu@mail.utoronto.ca
|