Floating Car Data: Travel Time Estimation in Urban Networks

Floating Car Data (FCD) is becoming a more and more popular technique for travel time measurements in road networks. Nevertheless, FCD is a sampling technique which requires controlling the statistical properties of link travel times to obtain accurate estimations. Based on microsimulation outputs, this paper shows which parameters play a key role in the travel time estimation accuracy, particularly in the case of urban networks. Among them, aggregation period and link definition are the most critical ones. They must be properly chosen according to the equipped vehicles ratio. INTRODUCTION Due to the vital need of accurate real time traffic data for supplying ITS applications, many a lot of research has been done, and is still ongoing in order to improve the efficiency of data acquisition techniques and data treatment methods. Among this data, travel time has been recognized to be one of the most valuable ones, particularly for ATIS (Advanced Traveller Information System) or ATMS (Advanced Traffic Management System). Travel time acquisition techniques are commonly divided into two groups: the indirect (e.g. deducted from loop detectors) and the direct acquisition (e.g. using AVI). Belonging to the second one, the Floating Car Data measurement is becoming more and more popular due, among other 1 reasons, to the mobile communication development and the fact that no road-based infrastructure is needed, reducing significantly maintenance cost and traffic disturbances. A wide range of studies has therefore been done on FCD based travel time estimation. The efficiency of this method is directly linked to the accuracy level with which estimated travel time calculated with sampled data (FCD) can match the “real” travel time experienced by the overall vehicles data set. Usually, travel time is calculated for each link of the network as being the average of the individual travel time recorded during the [t;t+∆t[ period, ∆t being the aggregation period. Consequently, the estimation error is generally defined as the difference between the sample and the full data set averages. Nevertheless, this accuracy indicator is not the most relevant as will be discussed in this paper. To be able to determine which estimation accuracy can be expected, individual travel time measurements evolution between and within the aggregation periods must be well identified. This evolution is significantly dependent on the type of network. Indeed, if the variability of a freeway link travel time data set (within an aggregated period) is generally low, it isn't the case for urban ones. Furthermore, bias problems have been highlighted (Sen & al. (1)) in urban network link travel time samples when the probe vehicles ratio isn't equal for the different turning flows leaving the link. Hellinga and Fu (2) have clearly demonstrated how the influence of traffic signals and platoon effects are basically responsible for these phenomena. The impact of this link travel time data set variability on optimal routes calculation has also been partially described by Sen & al. (4). Considering these observations an important reflection has to be made on the real sense of using only the aggregated travel time average to describe a data set with a wide variability. Indeed, Josias Ziestman and Laurence R. Rilett (3) have plainly shown the advantages of a disaggregated-based travel time estimation facing an aggregated one. Using field trial data allows only studying sampled travel times but not the corresponding full data set. Combining FCD measurements with AVI (Automatic Vehicle identification) on some links could be a solution but would be costly and limited to some individual and local applications. This is why this research relies on microsimulation outputs, using a well calibrated urban model of the Lausanne (Switzerland) downtown network. The latter has been built using the AIMSUN software (5) developed by the Polytechnic University of Catalunya, Spain. All the individual link travel time are recorded during the simulation process (a five hours period around the evening peak hour) and analysed to be able to describe the data set structure properties. This approach allows obtaining a totally disaggregated picture of link travel time measurements and not a limited one only based on averages. Relying on these microsimulation results, this paper underlines the specific travel time's 2 statistical properties for urban links and explains why decreasing the variability of data sets is essential to obtain a more accurate travel time estimation. Different techniques to achieve this goal are then suggested. The important problem of measurement lacks when the equipped vehicle ratio is too low is also described and implications on estimation preciseness are demonstrated. Different performance indicators for the travel time estimation accuracy assessment are then presented. Lacks of relevance induced by the use of the sample average to represent link travel times are also commented. Finally, further research orientations and conclusions are given. TRAVEL TIME VARIABILITY Figure 1 gives an example of individual link travel time records provided by the simulation for a major avenue of the downtown area. It shows an important data set variability which is typical for urban links. The red line represents the aggregated travel time averages calculated on the basis of 15 minutes aggregation periods. Figure 1: Individual and aggregated link travel time Urban link travel time variability is mainly due to two different phenomena. The first one is the medium-term evolution of traffic conditions. Indeed, changes in traffic flow according to the day time implies a change in queue length and in congestion level, those having a direct impact on travel time. This variability can easily be shown by the temporal evolution of the red line in Figure 1. It represents the general trend of the travel time evolution during the day. In the other hand, short-term variability is due to non continuous traffic conditions that vehicles face during their journey through the link. Traffic lights, stops and give way signs are the main causes of these flow disturbances. Short term variability can be deduced from the data included in-between the aggregation period limits. It represents the major difference 3 between urban and freeways or highways travel time. Indeed, vehicles usually find continuous traffic conditions during an aggregation period in the latter case. Short term variability plays an important role within the travel time estimation process as probe vehicle based link travel time estimation generally aims to match the aggregation period data set average with a limited sample one. It can be shown that the larger the data set variability, the larger the sample size (in percentage) must be to reach a predetermined matching level. This phenomenon highlights the importance of the short-term variability of link travel time. Consequently, to obtain a more accurate averages matching with a fixed probe vehicle percentage, the data set variability has to be reduced. Different techniques allow this variability to decrease. Among them, a more detailed definition of “link” is suggested. A “classical” link is indeed frequently defined as the arc linking two road intersections. Nevertheless, cars crossing a “classical” link controlled by traffic lights experience different traffic conditions according to the different turnings possibilities at the end of the link and also according to the link they are coming from. This is especially the case when the upstream junction is also equipped with traffic lights. By subdividing the “classical” link into sub-links, joining each possible link entrance with each exit, the variability of the corresponding data sets is, most of the time, smaller than the one with the “classical” link definition. An exploratory study made by Chen and Chien (6) has already underlined the necessity of distinguishing the travel time measurement according to the exit link, but for freeway networks only where the potential seems much lower than in urban ones. An example of how this sub-division is done is shown in Figure 2 and 3. Figure 2: Classical link representation