Traffic-monitoring systems, such as those using loop detectors, are prone to coverage gaps, arising from sensor noise, processing errors and transmission problems. Such gaps adversely affect the accuracy of Advanced Traveller Information Systems. This project will explore models based on historical data that can provide estimates to fill such gaps. We build on an initial study by Mr. Rafael J. Fernandez-Moctezuma, using both a linear model and an artificial neural network (ANN) trained on historical data to estimate values for reporting gaps. These initial models were 80% and 89% accurate, respectively, in estimating the correct speed range, and misclassifications were always between adjacent speed ranges (in paricular, the free-flow range and congested range were never confused). Going forward, we will investigate other non-linear models, such as Gaussian Mixtures, that provide further statistical metrics, in contrast to the uninterpreted weights of ANNs.
This work will exploit the Portland Transportation Archive Listing (PORTAL) at the Intelligent Transportation Systems Laboratory at PSU. Dr. Tufte helps supervise development of PORTAL, and Mr. Fernandez used PORTAL data in his study. PORTAL holds more than two years of Portland-area freeway-loop-detector data at both detailed and aggregated levels, and is an ideal resource for the proposed work.
Initially we will be building and testing estimators in off-line mode. We will select a highway segment (comprising multiple detector stations) that is representative in terms of pattern of outages. We will build models for this segment, then examine their performance on estimates for synthetic gaps (so we can compare estimates to reported values). Later, using live loop-detector data (which PORTAL supports), we will work towards on-line estimation over the local freeway network, which requires computing estimates in a timely manner. Our end target is improvements in end-user travel information products, such as the Portland-Metro Speed Map on ODOT's Trip Check.
Our main evaluation metric will be the trade-off curve bewteen accuracy of prediciton and percentage of gaps that can be filled.
Possible extensions to this work include:
(A) Surveying users of travel information products to assess the added value of gap filling and understanding of confidence levels.
(B) Detecting and adjusting unreliable data values, in addition to estimating missing values.
(C) Further evaluation by comparison to probe vehicle data.
(D) Investigating whether regimes with low-confidence estimates also see low reliability of travel-time estimates.
(E) Developing suites of estimators (for example, with different inputs) and combining their estimates, perhaps in a confidence-weighted manner.
This research supports national surface-transportation research priorities, including the Systems Management Information area (ITS JPO). Within that area, it relates to (2) Data Management (techniques and guidance for processing and managing data associated with highway and transit monitoring) and (5) Data Dissemination (exchanging information about transportation services and providing that information to travelers). [Page 3-15, U.S. Department of Transportation Research, Development, and Technology Plan, 6th Edition]
