Characterization of Congestion Based on Speed Distribution: Statistical Approach Using Gaussian Mixture Model

Traffic congestion has been one of major problems in modern cities and thus numerous measures have been taken to mitigate it. An appropriate selection of congestion mitigation strategies comes from the sound understanding of the congestion characteristics. Many researchers, thus, have tried to identify the effective ways for measuring level of congestion or roadway performance. As a result of these efforts, various approaches were developed for some specific purposes. It is, however, still necessary to develop new approaches to measuring congestion since congestion cannot be defined in a unique manner and there are numerous situations in which different types of congestion measure are required. This paper proposes a new approach to characterizing the congestion using Gaussian mixture model. This approach assumes that the speed distribution over a given time period has a form of mixed distribution which includes congested and uncongested ones. In addition, it supposes both the speed distributions are normal. Based on these assumptions, it was possible to easily identify the congestion characteristics by exploring the parameters of the distributions. Authors proposed the resulting parameters, means, variances and mixture weights are potential quantifying measures for the severity, variability and duration of congestion, respectively. In particular, this approach is easy to apply since the longitudinal speed data are the only input for the analysis. From the case study using one-week fourfreeway segments data, authors concluded that this approach might become a promising one in analyzing congestion characteristics, complementing the existing approaches. TRB 2005 Annual Meeting CD-ROM Paper revised from original submittal. Ko, J. and Guensler, R. 2 INTRODUCTION Background Traffic congestion has been one of major issues that most metropolises are facing and thus, many measures have been taken in order to mitigate congestion. It is believed that identification of congestion characteristics is the first step for such efforts since it is an essential guidance for selecting appropriate measures. Many researchers, thus, have tried to identify the intensity, duration and extent of the congestion using a variety of approaches. Stathopoulos and Karlaftis (1), for instance, tried to probabilistically model the duration of traffic congestion using loglogistic function. Thurgood (2) developed an index called Freeway Congestion Index, which simultaneously captures the extent and duration of congestion on freeways. In addition, some recent studies have introduced a reliability measure, so called “buffer time index”, which shows the effect of congestion on the reliability of travel rates along the roadway (3-4). Cottrell (5) developed logistic regression models with explanatory variables, AADT/capacity, K-factor (i.e., the ratio of the 30 highest hourly volume of the year to the AADT) to predict the occurrence of congestion. Vaziri (6) and Hamad and Kikuchi (7) applied fuzzy set theory, where multiple congestion measures were combined to get a single comprehensive measure. It should be admitted that the selection of adequate measures of traffic system performance is not an easy task. This is because congestion is rather subjective and at the same time a location-, facilityand time-dependent matter. Thus, congestion has not yet been clearly defined in a single manner. The congestion, however, can be generally defined as travel time or delay in excess of an agreed-upon norm. The agreed-upon norm may vary by type of transportation facility, geographic location, and time of day (8). Many other researchers have also pointed out the aforementioned aspects of congestion and proposed new approaches to quantifying congestion considering those factors. Congestion index should be developed and selected such that they can effectively and efficiently describe actual traffic conditions and contain useful information. Also, they must be understandable and acceptable to travelers and transportation experts and be well fitted to given purposes. To meet these conditions, the measures require appropriate data in terms of type (e.g. travel time, speed, and traffic volume), temporal, and spatial coverage, etc. Among others, travel time and speed based measures have been most widely used in previous research (8, 9). In fact, the measures associated with the time or speed are easy to understand and interpret. The current approaches to measuring roadway performance are well summarized in Medley and Demetsky (3). The target of congestion measures can be an area-wide network, corridors, or link segments. The different type of target requires different data collection efforts. Recently, Intelligent Transportation System (ITS) or Advanced Traffic Management System (ATMS) data are popular resources for transportation research since they are readily available and huge in quantity. Medley and Demetsky (3) is an example in which ITS information data were used to develop the corridor-level performance measures. Also, the GPS data from instrumented vehicles are growing as a popular data source for transportation studies. D'Este et al. (10) discussed the usefulness of GPS data when developing congestion indices. TRB 2005 Annual Meeting CD-ROM Paper revised from original submittal. Ko, J. and Guensler, R. 3