Using Archived ITS Data to Generate Improved Freeway Travel Time Estimates

Accurate travel time estimation has become possible with the deployment of advanced traveler management and information systems. Dynamic message signs, websites, and handheld/in-vehicle devices are being increasingly used by public agencies to communicate important travel information to the public such as incidents, road closures and travel times. Travel time estimates are usually derived from roadway sensors, although other technologies such as cell phone matching, license plate matching, automatic vehicle identification and video detection have also been employed. In Oregon, freeway travel time estimates are generated by the Oregon Department of Transportation using data from inductive loop detectors throughout the Portland metropolitan area. These estimates are generated using a simple midpoint algorithm that extrapolates measured speeds over a freeway segment. The objective of this paper is to evaluate the accuracy of the current midpoint algorithm by comparing the derived estimates to ground truth (probe vehicle) travel times. In addition, travel time estimates from a travel time algorithm developed by Coifman were also evaluated for accuracy under varying traffic conditions. Various scenarios were tested using traffic data from both upstream and downstream detector stations. The results indicate that both the midpoint and the Coifman algorithms generate accurate travel time estimates under free flow conditions. The Coifman algorithm using data from the upstream detector provided the best estimate of travel time for a link during congestion as well as in periods after an incident occurred. INTRODUCTION An important component of the management of the transportation system is the delivery of pre-trip or en-route travel information to the public in order to enable them to make informed travel choices. With reliable information, travelers can adjust their travel time, mode and route. This information should include travel time; it is relatively easy to report “current” travel time on instrumented freeways. It is more difficult to report arterial travel time and to forecast future travel times on network segments at times when travelers actually reach a particular segment. Thus, the accurate estimation of travel time has become critical with the advent, deployment, and maturation of advanced traveler information systems. Travel time information is delivered to the public through devices such as desktop PCs, handheld phones or PDAs, in vehicle displays, highway advisory radio or dynamic message signs. Travel time estimates are derived using techniques such as license plate matching, video imaging, cell phone tracking, automatic vehicle identification (AVI) as with electronic toll tags, automatic vehicle location (AVL) as with GPS tracking systems, and fixed sensors such as inductive loop detectors (1). Generating travel time estimates from fixed point sensors is the most common method in the U.S. Travel time estimates from dual loop detectors that directly measure speed are inherently more accurate because estimating travel time from single loop detectors involves making assumptions about vehicle length to calculate speed. Past research has proved that this assumption leads to higher inaccuracies (3,4). There are approximately 502 double loop detectors on the freeways in the Portland, Oregon metropolitan area. These detectors, located in each freeway lane and on 138 metered on-ramps, report count, speed and occupancy every 20 seconds and this information is transmitted to the Traffic Management Operations Center (TMOC). Detectors are spaced an average of 1.24 miles apart on the 140 directional miles of freeway in the region. A number of algorithms have been developed by researchers to estimate travel time from single loop detectors (5,6). The Oregon Department of Transportation (ODOT) provides real time traffic information through several dynamic message signs (DMS) at key decision points on the network and through the web (tripcheck.com). As shown in Figure 1, ODOT provides a range of travel time in minutes (e.g. 10-12 minutes) to a particular junction. Under congested conditions the time range increases, and under very congested conditions due to an incident, for example, the message will read INCIDENT. Travelers can also obtain vital information through their mobile phones by dialing 511. In order to generate real time travel time estimates on highways, ODOT employs a simple midpoint algorithm that extrapolates measured speeds over each freeway segment. The goal of this paper is to evaluate the accuracy of the midpoint algorithm and that developed by Coifman (2) compare the travel time estimates with ground truth (probe vehicle) estimates under varying traffic conditions. The overall objective of our travel time study is the identification and/or development of a travel time algorithm that is able to generate accurate travel time estimates especially during congested periods as well as during the occurrence of incidents. The remainder of this paper is divided into four sections. A brief literature review follows this section, followed by description of the data sources and the study site. Analysis of the travel time estimates produced by the different algorithms is described followed by conclusions and some recommendations. TRB 2007 Annual Meeting CD-ROM Paper revised from original submittal. Kothuri, Tufte, Ahn and Bertini 2 BACKGROUND Calculating speeds from loop detectors and extrapolating those speeds over a freeway segment to determine a segment travel time involves assumptions about vehicle length and calculating speed based on relationships between flow, occupancy and vehicle length. Past research by Hall and Persaud (3), and Pushkar, Hall and Acha-Daza (4) indicates that the accuracy of the speed estimates derived from an assumed value of vehicle length is poor, which in turn leads to less accurate travel time estimates. Recent studies have focused on improving the accuracy of estimated travel time from single loop detectors (5, 6). Dailey (5) used a stochastic approach and a Kalman filter to estimate speed and derived travel time using the speed estimate. However, this method requires estimation of several parameters. Petty et al. (6) assume that within a certain time frame, travel times follow the same probability distributions. Cortes et al.(1) derive travel times assuming that the representative travel time for a certain link is the travel time of the vehicle that reaches the midpoint of the link at the midpoint of the time interval. Coifman (2) postulates that a vehicle’s travel time across a link is the time taken for the vehicle’s trajectory to propagate across the link. The algorithm proposed by Coifman uses basic traffic flow principles to estimate the travel time of a trajectory across a link. The travel time is estimated knowing the vehicle’s velocity, the headway between the vehicles and congested wave velocity. This algorithm uses successive speed readings to build the trajectory of the vehicle in order to calculate travel time. Coifman’s algorithm has been tested with individual vehicle trajectory data but its usefulness in an environment with vehicle count and speed data aggregated over short time intervals (e.g., 20 or 30 seconds) was mentioned by Coifman (2), but has not been extensively tested. The midpoint algorithm used by ODOT is currently used to provide real time travel time estimates. The algorithm computes travel time as a ratio of distance to speed, using influence areas upstream and downstream of each detector. ODOT establishes influence areas at the midpoint between two detectors along a directional freeway segment. Due to the availability of a regional ITS data archive it was possible to design a customized user interface for testing various travel time algorithms. For this study, travel time estimates were generated using ODOT’s midpoint algorithm using archived ITS data were then compared to ground truth travel time data extracted from probe vehicle runs. (8) In addition, these travel time estimates were further compared with travel times generated by several versions of the algorithm developed by Coifman. STUDY AREA AND DATA In previous study, travel time estimates derived from the midpoint algorithm were compared to the probe travel times for the links shown in Figure 1 (8). This current study builds on the earlier work by evaluating additional algorithms in addition to the midpoint algorithm. A subset of nine links were chosen from the previous study to further evaluate the performance of ODOT’s standard midpoint algorithm and several variations of the one developed by Coifman. The summary of probe vehicle runs, including the link number, the length of each link in miles and the average detector spacing, and the number of probe vehicle runs for each analyzed link are shown in Table 1. Figure 1 shows the locations of the links. These links were chosen on the basis of availability of ground truth probe vehicle data and high traffic flows and congestion during peak periods. Travel time estimates were computed for Links 3, 4, 5, 6, 8, 9, 10, 12 and 13. Links 3, 4, 5, 6 covered portions of I-5, whereas links 8 and 9 represented OR-217. Links 10 and 12 represented I-205 and link 13 stood for I-84 E. TRB 2007 Annual Meeting CD-ROM Paper revised from original submittal. Kothuri, Tufte, Ahn and Bertini 3 FIGURE 1 Site map, influence areas for midpoint algorithm and ODOT VMS. Three data sources were used in this study. As part of the Portland region’s Advanced Traveler Managed Systems (ATMS), the ODOT Region 1 Traffic Management Operations Center (TMOC) maintains a fiber optic communication system linking all 502 inductive loop detectors. These detectors report count, occupancy, and speed in each freeway lane, and count from the on-ramps every 20 seconds. These data are fed into the Portland Oregon Regional Transportation Archive Listing (PORTAL—see http://portal.its.pdx.edu), a database that was developed at Portland State University based on the Archived Data User Service (ADUS) framework for archiving intelligent transportation systems data. PORTAL provides an extensive and valuable data set that can be used for improved performance assessment and modeling (7). A customized travel time functional area was set up in PORTAL to generate travel time estimates from archived loop detector data using the ODOT midpoint and the Coifman algorithms. The project would have been much more difficult without the PORTAL system. Travel time estimates generated from PORTAL’s archived loop data were compared with two sets of ground truth data. Probe vehicle data (87 runs) were collected during April–May 2005 for selected links of the Portland freeway network by researchers at Portland State University (8). A total of 15 hours of data, over 516 miles of travel over 7 days were collected. A statistical analysis was conducted to ensure that sufficient numbers of runs were performed. Typical probe vehicle headway ranged between 5-7 minutes. Travel time data for all the freeway TRB 2007 Annual Meeting CD-ROM Paper revised from original submittal. Kothuri, Tufte, Ahn and Bertini 4 links were collected using global positioning systems (GPS) devices. Custom software (ITS-GPS) developed specifically for use with Palm handheld computers and the GPS devices were used to record the position of each probe vehicle every 3 seconds. These data streams could also be used to calculate speed and distance traveled (9). In addition to the probe vehicle data from 2005, transit probe data was provided by TriMet for bus routes 95 and 96, which are express routes running on the freeway. Nine days of the northbound runs for route 96 were analyzed in this study. The route traced I-5 N between OR-217 and I-405 interchange can be represented by Link 3, shown in Figure 1. The TriMet buses are equipped with an AVL system that also archives detailed stop-level activities (10). For a three week experiment in November 2002, TriMet’s buses were programmed to record arrival times for “pseudo stops” located at fixed, designated points on the freeway, since buses do not stop on the freeway. The data from TriMet for these virtual detectors for northbound route 96 consisted of 148 runs and contained an arrival time and a leave time for each “pseudo stop” along with the distance traveled from the start to the end of the trip. Figure 2 shows the sample trajectories from Nov 7, 2002. These trajectories represented route 96, which traversed a section of I-5 N, just south of downtown Portland. The pseudo stops are indicated on each of the trajectories with a marker. FIGURE 2 Sample TriMet bus trajectories and Coifman algorithm travel time estimation. METHODOLOGY Travel time estimates from the archived loop detector data were generated using a standard midpoint algorithm and one developed by Coifman. The algorithm proposed by Coifman uses traffic flow theory to estimate travel times for a link (5). Coifman proposed that the velocity of a vehicle can be represented by: ( ) t u x f t x v . ) , ( + = (1) where x is the distance, t is time and u can be either uf the free flow signal velocity or uc the congested signal velocity. The vehicle trajectories in a time space diagram can be represented by the differential equation: ( ) t x v dt dx , = (2) The vehicle’s travel time across a link is the time taken by the corresponding trajectory to travel across the link and is shown in Figure 3. Travel time for a link can be estimated using vehicle velocity (vj), headway (hj) and congested signal speed using the relationship: