| The research effort in Advanced Traffic Management Systems (ATMS) focuses on developing more efficient real-time, adaptive traffic control systems to alleviate traffic congestion on urban roadways. In developing control strategies for signalized intersections, predicted traffic demands on each turning movement are essential inputs for signal timing plans. The most commonly-used representation of these traffic demands on a bounded network of intersections consists of a specification of (1) the boundary arrival flows at each entry point of network, and (2) the turning movement proportions at each approach of intersections. Coupled with a well-supported dynamic traffic flow model, this knowledge allows one to predict turning volumes at each intersection in the network. The under-determined methods initialized in the late 1970's for estimating turning volumes are subject to the criticism that none has been shown to yield consistent estimates. The over-determined methods, based on the least-square criterion that minimizes measure of discrepancy between observed and predicted time-series output counts, requires that a full set of detector counts be totally available. Such a dense detector configuration imposes prohibitive cost on installation and maintenance, which is not economically feasible in modern traffic surveillance systems. The ordinary least-square methods also fail locally whenever detectors are ill-functioned. These motivate us to develop system identification algorithms in both off-line and recursive forms, based on nonlinear least-square (NLS) and quasi-maximum likelihood (QML) criteria, to produce consistent and efficient estimates of turning movement proportions under the situation where only partial set of detector counts is available. |
