Identifying the Onset of Congestion Rapidly with Existing Traffic Detectors

Problem

Traffic surveillance systems need to reliably identify the presence of congestion, and to respond rapidly when congestion is detected. Loop detectors are good at the first, but not so good at the second task. This paper presents an algorithm using existing dual loop detectors in a new way to improve their performance on these tasks; it promises to be deployable in the short term

Method and Findings

Loop detectors calculate flow and occupancy at discrete points on a freeway. We generally assume that this information represents the space between links as well. However, under congested conditions this is not always so; the traffic stream can change significantly from moment to moment in both directions, especially when traffic is heavy.

Conventional loop detectors easily identify the presence of congestion, but response time is usually slow. Our method solves this by looking for free-flow traffic instead of congestion. Recognizing that signals emanate from an incident in two directions-a drop in speed moving backwards at about 8 mph, and a drop in flow moving forward at the prevailing traffic speed-the algorithm looks for a fast-moving drop in flow, thus quickly identifying the onset of delay.

The drop in flow is found by identifying long vehicles at a detector station, then looking for a similar vehicle in the same lane upstream within a certain time window, defined by the amount of time it would take the vehicle to travel that far at free-flow speeds. If the vehicle is identified, then the presence of free-flow traffic reveals itself. Since lane changes are less common under free-flow conditions than congestion, the likelihood of identifying the vehicle is higher. Using long vehicles, which can be four times as long as smaller vehicles and thus easier to identify with some certainty, we can quickly ascertain whether free-flow conditions are met, or not.

This work is compatible with existing detector infrastructure, and simple enough to implement on existing Model 170 controllers, which are based on twenty-year-old computer technology. It is intended to augment, not supplant, conventional surveillance strategies to improve performance.