You are on page 1of 10

The Potential of Microscopic Simulation in Traffic Safety and Conflict Studies

Iisakki Kosonen Royal Institute of Technology, Centre for Traffic Simulation Research 100 44 Stockholm, Sweden Helsinki University of Technology, Transportation Engineering P.O.Box 2100, FIN-02015 HUT iisakki.kosonen@hut.fi Suhan Ree Los Alamos National Laboratory, Simulation Applications Group MS M997, Los Alamos, NM 87544, USA suhan@lanl.gov

BACKGROUND Traffic safety is a major problem in road traffic systems. Although the problems are serious it is difficult to study them systematically due to the fact the number of actual collisions is very small compared to the whole traffic flow and potential accident cases (conflicts). There is a growing interest to use computer simulation in order to tackle the problem systematically. Microscopic traffic simulation has been successfully used in evaluation of traffic efficiency and emissions, but not yet widely in studies of traffic safety aspects. This paper will discuss the potential of traffic safety simulation based on experiences from simulation studies in two research laboratories that were using very different types of traffic simulators. TRANSIMS (Transportation Analysis and Simulation System) is a high-speed parallel micro simulator that allows simulation of large metropolitan areas on microscopic level. TRANSIMS was developed by the Los Alamos National Laboratory in New Mexico, USA. It utilizes synthetic populations to create the traffic demand and cellular automata for high-speed parallel simulation of traffic flows (Smith et al. 1995). HUTSIM is a very different kind of microscopic simulator developed by the Helsinki University of Technology, Finland. The object-oriented HUTSIM model is built for very high precision simulation of interactions between vehicles, pedestrians, traffic environment and traffic control (figure 1). Because the high fidelity of modeling requires lot of computing power, very large networks cannot be simulated. A common feature of TRANSIMS and HUTSIM is that they both use rule-based vehicle dynamics. In HUTSIM a set of few rules controls the selection of discrete speed levels and lanes (Kosonen 1999). In TRANSIMS also the space is discrete which leads to a very minimal and hence fast rule sets for control of speed and lane changing (Simon Nagel 1998).

Figure 1. High-fidelity modeling of interactions with the object-oriented HUTSIM-simulator.

MODELING TRAFFIC SAFETY Using simulation for traffic safety studies is a complicated matter. Since the simulated traffic is usually collision free, the safety aspect is already built in to the model. In other words the simulated driving dynamics is derived from safety requirements and without any exceptions no accidents or even conflict situations will occur. Two types of solutions can be used to allow safety studies with initially collision free simulator. One is to introduce some erroneous behavior to allow conflicts and accidents to take place. This approach has been used with HUTSIM as explained later. The other approach is to take the collision free driving dynamics as input to the safety model, which then assigns some probability of accident to each potentially risky driving situation. This type of approach was tested with the TRANSIMS system (Ree et al. 2000). There are different levels of fidelity in modeling traffic systems. In macroscopic modeling no individual cars are identified and traffic is considered as a continuous flow. Therefore macroscopic simulation is not suitable for safety modeling and conflict studies between individual cars. With low fidelity microscopic simulators (like TRANSIMS), it is possible to model safety issues in rough level. The output of cellular automata must be postprocessed and filtered in order to achieve more realistic car behavior. High-fidelity microscopic models (like HUTSIM) offer plenty of possibilities for systematic examination of conflict situations between individual vehicles.

The most suitable model for traffic safety simulation is a so called nanoscopic model. This means very detailed modeling of driver perception, decision making and actions including possible limitations and errors within them (figure 2). A nanoscopic model gets very complicated, which requires a lot of computing power. This means that it is not possible to simulate all metropolitan traffic using nanoscopic simulator, as it is possible with low-fidelity models such as TRANSIMS. However, a nanoscopic model can also be used to feed driver behavior data to a micro simulator that runs faster.

Figure 2. Nanoscopic modeling of driver behavior including various types of errors. There are numerous factors that affect the driving performance, like age, experience, attitude, state of mind, alcohol/drugs etc. Various factors are discussed for example by Ledoux & Archer (1999). It is very difficult to include all these factors into a simulation model. However, if we assume a certain level and distribution of driver errors and reaction time, it is possible to simulate the consequences, i.e. what kind of conflict situations arise. It is also possible to examine how traffic flow, speed limit, traffic control, sight distances and complexity of the traffic environment affect the traffic safety. It is also most interesting to study the effects of ITS on traffic safety, especially when only part of the vehicles is equipped with certain devices while others are not (i.e. the effect of penetration rate of ITS-devices). It is essential that the simulation model includes sufficient variance in the driver behavior. To produce risk situations some of the drivers must occasionally be inattentive, sleepy or disoriented. In micro simulation these things can be modeled as temporary lack of observation (short blackouts) and hence increased reaction times. The calibration and validation of such a simulation model is not an easy task, since many of the behavioral factors cannot be directly measured. One must especially study the real conflict situations and try to produce similar situations in the simulation model.

CONFLICT STUDIES USING HUTSIM Recently a lot of interest has arised towards usage of micro simulation for traffic safety studies. This has led to several projects that are exploring the potential of simulation in traffic safety research. A research project called SINDI (Safety INDIcators) was started in Sweden by the Royal Institute of Technology. The main objective is to produce reliable indicators of traffic safety especially for evaluation of the effects of ITS. A new simulator will be built based on HUTSIM micro simulator developed by Helsinki University of Technology. Similar project is also planned between Helsinki University of Technology and the Technical Research Centre of Finland. New features and concepts are needed in the HUTSIM model in order to make it suitable for safety studies. Originally HUTSIM was developed for simulation of traffic in signalized intersections. The traffic model was then very deterministic producing very homogenous driving behavior. When HUTSIM was enhanced towards motorway simulation, more variation and distributions were added into the driving behavior. These features are also needed in the safety simulation since accidents and conflicts typically take place in exceptional situations i.e. in the tails of the distributions. Typical distributions needed are the distributions of the desired speed and reaction time as well as the minimum accepted gaps in car-following, lane changing and yielding. In the HUTSIM model the drivers are characterized as calm or aggressive drivers between 0-99. The calm drivers tend to choose the low end of speed distributions and the high end of the gap distributions, while aggressive drivers tend to choose the opposite i.e. high speeds and short gaps. This arrangement gives sufficient variation in driving behavior while keeping consistency between adjacent choices of individual drivers. Reaction time is an essential feature in traffic safety simulation. Typically in simulation models the reaction time is a constant value obtained by updating the vehicles in a certain order. In HUTSIM no updating order is presumed and the delay between an external event (like a pedestrian suddenly in front) and driver reaction is obtained in an other way. A human driver can react fast only if the emerging event is currently in his focus. However, usually the driver is sharing his focus with other things by scanning the mirrors, car equipment, traffic environment and traffic signs. The random delay caused by this scanning is modeled in HUTSIM by updating the driver perception/decision more rarely than the rest of the simulation model. Since the whole simulation model is typically updated every 0.1 second, any event that the driver should react to, can take place every so often. However if the driver decision is updated say once per second, then inherently a random reaction delay is generated between 0.0-1.0 seconds. In this models it is assumed that immediately after the first reaction to the external event, the driver turns to a attentive state where his perception and decisions are updated as often as the rest of the simulation model. However, after a given time of no events, the driver returns back to the relaxed mode where his reactions are delayed. To elaborate this model, the driver perception frequency could depend on the driver type. A successful safety and conflict simulation requires modeling of two-way interactions between simulation agents like vehicle/drivers and pedestrians/cyclists. In the basic HUTSIM model yielding in intersections is performed by yield signs that show stop-signal when the gap is too small and go-signal when the gap is large enough. For each situation the critical accepted gap can be obtained from a distribution. However, in this model the vehicle with right of way never

reacts to the yielding vehicle (they have always go-signal). A modification was made recently to allow prioritized vehicles to react (decelerate) if the yielding vehicle is choosing a too small gap and hence causing a conflict situation. These conflict situations are detected and saved into data files for later analysis. According to conflict analysis (Hydn 1987), the conflict situations can be divided into two portions (serious/non-serious) by experimentally studying the situations where drivers want to be and where not. Areas of serious/non-serious conflicts can be divided by a curved line in space of time to collision and speed. Also in simulation the speed and time to collision (ttc) can be registered at the first moment when the vehicle has to brake. This way a similar plot can be created that was obtained by the field studies. Also the severity of the simulated conflicts can then be classified based on experimental studies (curved line). Also the whole trajectory of vehicle deceleration can be demonstrated in the same way (figure 3.). This kind of approach is the first step towards systematic research of conflicts by using simulation model. If the number of conflicts in a given intersection can be evaluated, then it is quite simple to calculate an estimation of the number of accidents in the same intersection.

Figure 3. Simulation of conlicts in a simple intersection (Y=Speed, X=TTC). A modified HUTSIM simulation model has already been used for preliminary safety studies. In the city of Helsinki the safety of pedestrian crossings has been studied by field measurements and by simulations. The simulation results confirmed the conclusion that reducing the speed limit in some critical intersections with high pedestrian volumes has very significant impact on the amount and severity of pedestrian accidents that are usually caused by a free vehicle.

HUTSIM has also been used for evaluation of some ITS-functions like ISA (Intelligent Speed Adaptation) and AICC (Automatic Intelligent Cruise Control). The simulations were made using a model from the ARENA-test site in Gothenburg, Sweden. The combination of ISA and AICC significantly reduced the number of severe pedestrian accidents as shown in figure 4. These early trials demonstrate the potential of microscopic simulation in safety studies. However, as stated earlier it is important to elaborate the driver model with features that take in account errors in driver perception and reactions. The target is to combine HUTSIM with detailed nanoscopic driver model.

Figure 4. Simulation of pedestrian accidents with various penetration rate of ISA-control.

ESTIMATING ACCIDENTS WITH TRANSIMS In high speed simulation of large networks it is not possible to model the driver behavior in a very detailed way. In that case, rather than modeling accurately the driving dynamics and conflict situations, it is possible to use a low fidelity model and methods for estimating the accident probabilities. A more detailed presentation of the method discussed here is written by Ree et al., (2000). The TRANSIMS traffic model is based on cellular automata (CA). In this model vehicles jump from one cell to another (7.5 meters long), hence the vehicle speed is always N*7.5 m/s. The speed levels are controlled by a set of rules. Another set of CA-rules is used to control lane changing and yielding in intersections (Nagel Schreckenberg 1992).

The CA-model produces rather rough speed trajectories, which have to be pre-processed for safety analysis (figure 5). TRANSIMS provides one snapshot output for every update. Therefore the speed and position of each vehicle can be obtained for every second. First the output file must be sorted to identify the individual speed trajectories. Then the CA speed trajectories must be low-pass filtered to produce more realistic driving behavior.

Figure 5. Computing accident likelihood from TRANSIMS simulation output. The next step is to find the deceleration events that can be potentially risky. For each deceleration event the required braking power is calculated (required to avoid collision). Also the time to collision can be calculated at the beginning of braking. It is assumed that these two factors can be used to estimate the probability of accident of any given braking situation. The individual probabilities can be aggregated for each link over a given time interval like 15 minutes. These accident probabilities can be demonstrated with the TRANSIMS 3D visualizer. In figure 6 the development of accident probabilities are shown in a test network. The height of the bars indicates the accident probability while the colors are set according to the traffic densities. After estimating the accident probabilities in each link, one can perform a monte-carlo simulation (roll the dice) to determine whether an accident really took place in any given link at given time interval. Also the actaul accidents can be demonstrated by the 3D visualizer. Finally after the accidents have been generated a new microsimulation can be started where accident are actually generated. This way the effects of accidents to the city traffic can be studied. Obviously external measurement data is needed to estimate the accident probability. Here a cumulative distribution of deceleration power was used. In this distribution low and moderate braking powers are most common. It is assumed that the rare events of high braking power are closely related to the high risk of accidents. It seems to be plausible that the low probability of braking power (rare events) correlates with high probability of accident. In accident simulation we are interested only in the rare events i.e. the tail of the distribution. We can assume that with low and moderate braking power the probability is close to zero. However, with high braking power we assume that Pcoll=C(1-Pdec), where Pcoll is the probability of accident for a braking

event, Pdec is cumulative probability of that braking power to take place and C is a constant value much less than one. The basic assumption has not yet been fully proved, but we believe that the method is already consistent enough to provide reliable relative safety indicators. By tuning the parameters to match with actual number of accidents, the model will probably also predict reliable absolute number of accidents. So far the parameters are tuned to produce similar braking power distribution than the measured one, which is a good start.

Figure 6. Demonstating the simulated accident probabilities (height=probability, color=density).

CONCLUSIONS AND FURTHER RESEARCH Traffic safety simulation certainly seems to offer a lot of potential. However, after the first trials it is still early to state if it is going to be a real breakthrough in traffic safety research. It can be said for sure that microscopic simulation offers a good platform for systematic studies of conflict and accident situations. It is also very likely that good relative safety indicators can be produced to compare different traffic arrangements and environments. This includes also the assessment ITS with various types of in-vehicle equipment and each type with various penetration rates.

Safety related types of ITS that can be assessed with simulation include at least Intelligent Speed Adaptation (ISA), Automatic Intelligent Cruise Control (AICC) and collision warning systems. Low-fidelity simulation combined with a separate accident estimation module seems to be relatively simple and efficient method for computing safety indicators. However, this method is likely to ignore many important factors that affect traffic safety, like driver perception and errors. Nanoscopic modeling of driver behavior, perception and errors may offer an interesting research platform for testing various hypotheses of driver behavior. However, due to the very complex nature of the model it is unlikely that this type of model is going to be an everyday traffic planning tool very soon. It also depends on how fast the price of computing power is going down, since nanoscopic modeling in larger scale requires much more computing power than the present microscopic models. A lot of further work is needed to enhance the HUTSIM simulation model for traffic safety assessment. In principle the present model can already be used for basic conflict studies by adding appropriate outputs for that purpose. Simulation of ITS from safety point of view requires modeling of the driver together with the supporting equipment. A nanoscopic driver model can be attached to the microsimulator as additional module that replaces the default vehicle control. It is also important to do intensive research of driving behavior especially in conflict situations to obtain realistic rules and distributions for the driver model.

REFERENCES Hydn C (1987), The development of a method for traffic safety evaluation: The Swedish Traffic Conflicts Technique. Dept. of Traffic Planning and Engineering, Lund University, Lund, Sweden. Kosonen I (1999). HUTSIM - Urban Traffic Simulation and Control Model: Principles and Applications. Doctoral thesis. Helsinki University of Technology, Transportation Engineering. 248 p. Ledoux C, Archer J (1999). Basic concepts of road safety. Report of the SINDI-project. Royal Institute of Technology, Centre for Traffic Simulation Research. Nagel K, Schreckenberg M (1992). A cellular automaton model for freeway traffic. Journal de physique I. pp. 2221-2229. Ree S, Kosonen I, Eubank S, Barrett C (2000). Estimating accidents using TRANSIMS. Paper submitted for TRB annual conference 2001. LA-UR:00-681. 15p. Simon P, Nagel K (1998). Simplified cellular automaton model for city traffic, Physical Review E, 58, 1286-1295. Smith L, Beckman R, Anson D, Nagel K, Williams M (1995). TRANSIMS:TRansportation ANalysis and SIMulation System. Proceedings of 5th National Transportation Planning Methods Applications Conference, Seattle, WA, pp.

You might also like