A Probabilistic Queue Prediction Model Using a Stop line Detector

Abstract

Traffic management systems rely on traffic signals, sensors, and skilled engineers to fine-tune the system. In New Zealand, Australia and some other countries, SCATS (Sydney Coordinated Adaptive Traffic System) is the only area-wide traffic control system deployed. One of the longstanding shortcomings in SCATS is that it does not have any knowledge of the queue length. This requires on-going manual calibration of the offset plans using on-site observations of the average queues. This research sets out to develop a queue estimation model using only the stop line sensors. The model is calibrated and validated against data collected from four signalised intersections in Auckland, publicly available NGSIM data and simulated data using VISSIM. A simulation-based case study is conducted in Auckland using SATURN traffic model to evaluate the extent of efficiency gains that can be achieved. A better estimation of queue length enables more efficient allocation of green times. A historical database of queue length is also useful for offline analysis and optimisation when fine-tuning traffic control systems without a queue estimation algorithm. A new logistic regression model developed in this research can predict queues of up to 20 vehicles with an accuracy of 83% to 93% with an MAE of fewer than two vehicles. The new two-regime queue discharge method shows that the current queue discharge methods underestimate saturation flows by up to 100 vph. A low-cost chase car GPS data collection method used in this research showed a much lower acceleration and deceleration rates than the one suggested in the AUSTROADS traffic signal guides and the popular SIDRA intersection model. The case study showed that substantial efficiency gains are possible using a unified queue estimation, offset and phase split algorithm and cycle-by-cycle optimisation method. Travel time savings were in the order of 20% over the existing optimised SCATS signal settings. An improved estimate of queues and delay can be useful to influence driver routing