Dynamic Origin-Destination Demand Estimation From Link Counts, Cellular Data And Travel Time Data

Abstract

Accurate real-time Origin-Destination (O-D) demand estimation is critical for development of Advanced Traveller Information Systems (ATIS) as well as Advanced Traffic Management Systems (ATMS). Traditional O-D estimation techniques typically pose it as optimization problems involving offline calibration of O-D priors from historical data and updating these prior based on real-time data sources such as link counts, link travel times and Global Positioning System (GPS) probe data. However, two major drawbacks have emerged from this approach. Firstly, the generation of effective priors requires extensive population data with high temporal sampling frequency. Such data is generally hard to obtain and process with low memory and compute power. Secondly, the optimization problems proposed for are often non-convex in nature. Therefore, convergence to an optimal solution is not guaranteed. Moreover, the algorithms suggested don’t scale well with increase in the size of the transportation network. In order to tackle these challenges, we first generate priors through a convex optimization framework by Wu et al. (2015) for O-D estimation in quasi-static settings from cell phone Call Detail Records (CDRs). We extend this approach to a completely dynamic setting by utilizing real-time link counts and link travel times through an Entropy Maximization framework proposed by Janson et al. (1992). Approximate solution to the traffic assignment sub-problem is also achieved from Wu et al. (2015). We test this procedure for estimating commute demand for a simplified freeway network representing nine counties of the San Francisco Bay Area. We estimate demands for 2,916 O-D pairs for three hours at five-minute intervals, making it a 104,976 dimensional problem. We analyze the spatio-temporal distribution of errors, the effects of wider cell-phone coverage and updating estimates over time. Results indicate that an increase in cell-phone coverage from 0% to 100% leads to a reduction in average Root Mean Squared Error (RMSE) from 15.7 to 7.2 while updating estimates also leads to reductions in RMSE when O-D demand is high.