Development and evaluation of potential accident scenarios involving pedestrians and AEB-equipped vehicles to demonstrate the efficiency of an enhanced open-source simulation framework.

This study introduces a method that allows the generation and safety evaluation of a scenario catalog derived from potential car-pedestrian conflict situations. It is based on open-source software components and uses the road layout standard OpenDRIVE to derive participants' motion profiles with the support of available accident data. The method was implemented upon the open-source framework openPASS and can simulate results for different active safety system setups and facilitates the prediction of system capabilities to decrease the relative impact velocities and collision configurations such as the point of impact. A demonstration case was performed where the scenario catalog was derived and used to evaluate pedestrian collisions with and without a generic autonomous emergency braking (AEB) system. The AEB system aims to intervene in the event of an impending collision and might affect the outcome of a baseline scenario. The study indicated a change in the collision configuration and identified conflict situations less affected by the system. A particularly interesting finding was that some scenarios even led to a higher number of collisions (at lower impact) for the AEB intervention in comparison to the baseline cases.


Introduction
In 2016, 22% of all killed European road users were pedestrians (European Commission, 2018). Effective protection of this vulnerable group is thus one of the major challenges for the overall reduction of fatalities. It is assumed that the market penetration of new active safety systems, pushed forward among others by regulations (Cieślik et al., 2019) and consumer tests (Euro NCAP, 2019), will decrease the number of (pedestrian) accidents (Rosén and Sander, 2009;Luttenberger et al., 2014;Kalra and Groves, 2017), reduce impact speeds and therefore the risk of injuries (Lindman, Ö dblom et al., 2010;Wisch et al., 2013;Jeppsson et al., 2018). Such investigations are often done virtually by counterfactual (what-if) simulations based on reconstructed real-world accidents (Lindman, Ö dblom et al., 2010;Rosén et al., 2010;Hummel et al., 2011;Jeppsson et al., 2018;Gruber et al., 2019). In comparison to physical tests (as performed in Vertal and Steffan, 2016;Kovaceva et al., 2020), simulations are cost-efficient for comprehensive system evaluations and compliment assessments of the efficacy and potential safety benefits of system parameters (Hummel et al., 2011;Hamdane et al., 2015;Gruber et al., 2019).
The overall objective of this study is to introduce a conceptual simulation framework to evaluate the efficacy of active safety systems. Section 2 emphasizes the design decision of the framework to rely on open standards and open-source software components. The introduced method in Section 3 should reveal how a multitude of potentially dangerous scenarios, involving pedestrians and AEB equipped cars, can be automatically generated, simulated, and evaluated in a transparent and comprehensible manner. The results of a demonstration example of the framework are shown in Section 4. Section 5 addresses the advantages and limitations of the introduced framework as well as further applications and possible extensions in the future.

Background
The virtual representation of traffic scenarios is firmly connected to the capabilities of the simulation environment in which they are modeled and analyzed. A scenario catalog for the efficacy assessments of active safety systems requires a proper representation to be implemented into the respective simulation environment. For the implemented method in Section 3, it is necessary to examine design decisions on the selected simulation environment and on the generation of the scenario catalog.

Simulation environment
Real-world accident scenarios are often analyzed in environments which allow modeling vehicle dynamics before and after a crash (Datentechnik, 2016). The efficacy of active safety systems for those scenarios is further assessed with additional tools such as those introduced by Wille and Zatloukal (2012), Seiniger et al. (2013) and Kolk et al. (2016). Published results with regard to the efficacy of active safety systems vary, however, by author, making it difficult to draw conclusions about the effect on future accidents. Predictions for the potential of avoiding accidents range from 20% (Lindman, Ö dblom et al., 2010;Rosén et al., 2010;Hummel et al., 2011) to up to 50% and more (Hamdane et al., 2015;Gruber et al., 2019). Besides the difference in data sources and considered baseline scenarios, this inconsistency possibly arises due to the variations in simulation setups and environments among the different studies (Wimmer et al., 2019). The harmonization group Prospective Effectiveness Assessment for Road Safety (P. E.A.R.S.) attempts to cope with this issue and has defined requirements for simulation platforms (Page et al., 2015). According to their suggestions, system models, including vehicle dynamics, sensors, and active safety systems, which might influence each other, are needed. Further, they suggest incorporating environmental influences, i.e., weather conditions, other road users, and traffic rules. Based on the requirements by P.E.A.R.S., the open-source initiative openPASS (Dobberstein et al., 2017;Wang et al., 2018) attempts to provide a transparent simulation platform for the efficacy assessment of active safety systems, which makes it a promising option.

Scenario catalog
According to the definitions in Ulbrich et al. (2015), a scenario can be understood as the temporal development of a specific event (such as a crash) caused by the road user actions and can be described with different levels of abstraction (Menzel et al., 2018). This reaches from functional descriptions, containing road layouts, traffic participants and their intention to concrete realizations, which contain, amongst others, trajectory information. According to the definition by Menzel et al. (2018) real-world accident scenarios, on which many counterfactual simulations are based, can be classified as concrete and predominantly follow the Pre-Crash-Matrix (PCM) (Schubert et al., 2012) convention.
Using reconstructed accidents as a data source for efficacy studies clearly benefits from incorporating extreme and rare situations. There are, however, drawbacks and shortcomings. According to Menzel et al. (2018), testing concrete scenarios requires a larger amount of scenarios to draw conclusions about any safety benefits of the assessed system. The number of available reconstructed accidents is however limited.
An open, holistic approach to generate a catalog of critical virtual testing scenarios of car-pedestrian conflicts would be a benefit for the efficacy assessment of active pedestrian safety systems. In order to derive such a catalog, it is essential to incorporate the influential factors central to the scenario generation process. Lindman et al. (2011) introduced a clustering scheme for conflict situations based on the intention of the drivers and pedestrians in motion patterns, e.g., Straight Crossing Path, Pedestrian from Right. To the best knowledge of the authors, the classification of conflict situations is currently only used for the statistical evaluation of accidents, it has had limited applications for the generation of a scenario catalog (Nitsche et al., 2018). Conflict situations however can be studied to determine how the road layout of the accident site and participant intentions, when combined, allow road users to reach a specific target destination within the road network. Sophisticated road network descriptions are required in a variety of applications ranging from traffic planning and traffic flow simulation (Cameron and Duncan, 1996;Fellendorf and Vortisch, 2010;Lopez et al., 2018), driving simulators (Dosovitskiy et al., 2017), to the realization of Advanced Driver Assistance System (ADAS) (Bender et al., 2014). Depending on the application, the formats differ in their granularity. Driving simulators often rely on very sophisticated and detailed descriptions, such as RoadXML (Chaplier et al., 2010) and OpenDRIVE (Dupuis et al., 2010). Modelling conflict situations through road layouts could therefore be of additional value to derive scenario catalogs, which can be used for the efficacy assessment of active safety systems.

Method
The method section introduces the framework's components, which form a toolchain as shown in Fig. 1. The output of one component is used as input for the subsequent one. The toolchain also describes the outline of this section and the shading in Fig. 1 identifies the contributions of this study to the state-of-the-art.

Simulation platform
OpenPASS is a novel simulation platform mainly developed for the efficacy assessment of active safety systems. It is capable of modeling the dynamic behavior of traffic participants, which consist of vehicle dynamics, sensors, and active safety systems. The simulation logic ensures a quasi-parallel simulation of all traffic participants, also referred to as agents and facilitates dynamic interactions. To simulate a particular scenario, scenario information as well as the algorithmic descriptions of the agents' systems must be passed to the platform.
The scenario input files contain the intended trajectory of the participants defined as PCM target trajectories. Target trajectories are time dependent discrete 2D target positions p i = (x i ,ŷ i ) of all traffic participants with a constant timestep Δt = 0.010 s. In addition to target trajectories, static objects such as buildings or lane marks can be defined. Agent system descriptions are hierarchically divided into components and connections. Connections enable communication and data exchange between components. The main elements of a component are library, schedule, and parameters, where the library points to the binary algorithmic description, which is dynamically linked with the component of other agents in the simulator. The schedule indicates the frequency with which a specific component is executed and parameters allow component configuration.
The simulated output is written to a separate file for each agent and contains the temporal information of the simulated scenario, comprising each agent's simulated positions p i (x i , y i ) in the global coordinate system, its velocities v i , and yaw angles ψ i .

Simulation models
OpenPASS contains a parameterizable two-track vehicle agent, which was extended with a generic AEB component for this study. Further, a pedestrian model was defined by the needs of the simulation. In the following sections, the properties of both models are shortly outlined and described.

Car agent and generic AEB component
The standard car agent of openPASS consists on the one hand of a mechanical two-track vehicle component (Datentechnik, 2016), which incorporates a TMeasy tire model (Hirschberg et al., 2007) and on the other hand a PID-controller based trajectory follower component. The second is required to cope with possible abrupt changes in the target trajectory, which are unfeasible due to the physical constraints. Therefore the PID control algorithm calculates a steady steering u δ , braking u B and throttle u T signal by incorporating the error between the nominal target positions p i (x i ,ŷ i ), the current position p i (x i , y i ) and future trajectory points.
The AEB extended car model, developed for this work, is outlined in Fig. 2. The objective of the AEB is to intervene in the event of imminent conflicts and consists of a sensor model based on Kolk et al. (2016), and an updated braking strategy (Gruber et al., 2019). The braking strategy is implemented in the Algorithm Selector model, which forwards the control signals to the vehicle dynamics model.
• Sensor: Perceiving the agent's environment is one of the prerequisites of the AEB for deciding whether or not an emergency brake intervention should be initiated. Hence, an idealistic sensor model, described in Kolk et al. (2016), was implemented. The sensor is mainly responsible for the perception of the environment in which it is moving. To do this, the sensor model scans a segment of a circle while transmitting detection rays to detect surrounding objects. The sensor field of view is shown in Fig. 3. Ray intersection points with the adjacent road user O = o 1 , …, o n act as origins for reflected rays. The reflections are in the direction of the relative velocity v r → , which is calculated by the ego velocity v e → and the respective velocity of the pedestrian v p → given in Eq.
(1). Using these rays, collision points, denoted as C = c 1 , …, c n , between and the ego car's front and the adjacent road user can be calculated. The minimal distance s min between these intersection points is used to calculate the time to collision (TTC), which is described in Eq. (2). The TTC is further used as a metric to trigger the brake signal of the generic AEB system. Determining the velocity v p → is a complex task requiring detection and tracking algorithms, which are highly dependent on the underlying system to perceive the environment. For this investigation, v p → was directly retrieved from the simulation environment. An adjacent road user is deemed to be recognized if it is intersected by one or multiple sensor rays. In comparison to real AEB implementations, the representation of the sensor as optimal, meaning that the time for data acquisition and object classification is instantaneous. • Braking strategy: The original car model, provided by the open-PASS, did not incorporate an actuator/brake delay and this has been implemented in this study by the authors. The delay is an additional input parameter of the generic AEB system which can be configured as required. The same holds for the brake signal, which is triggered when the TTC is below a specific configurable threshold t trigger . The build-up time for the brake has been modeled as described in Gruber et al. (2019). The braking system response is depicted in Fig. 4 and consists of a brake delay and the build-up time until the full brake.
The developed toolchain to generate and evaluate of a scenario catalog derived from potential car-pedestrian conflict situations. The framework is comprised by different components, based on open standards and frameworks. Applied, extended and novel components are highlighted through different shadings.

Fig. 2.
Model of the AEB equipped car as used for the current study. The system model contains the car components and the channels for data exchange.

Pedestrian
The public available version of openPASS does not contain a pedestrian agent, which made its implementation necessary for this investigation. The movement of pedestrians in road traffic is in many ways complex. Despite the great acceptance of theoretical models for describing group phenomena such as done by Helbing and Molnár (1995), accurate statements about individuals movements are difficult. Pedestrian movements are strongly influenced by intrinsic factors such as age (Bohannon, 1997) or extrinsic factors like traffic rules (Fugger et al., 2001) or the surrounding traffic (Yannis et al., 2013). In extreme situations, such as in an emerging accident, the behavior further differs from the ordinary and pedestrians show avoiding reactions as investigated by Soni et al. (2013) and Schachner et al. (2020). Current studies on the efficacy of active pedestrian protection systems simplify pedestrian behavior. This concerns real tests (Euro NCAP, 2019) as well as simulations (Lindman, Ö dblom et al., 2010;Rosén et al., 2010;Hamdane et al., 2015;Jeppsson et al., 2018;Gruber et al., 2019) of reconstructed real-world accident scenarios. The speed of pedestrians is usually assumed to be constant and pedestrian models do usually not react to the approaching vehicle. Stochastic modeling of pedestrian behavior as shown in Huang et al. (2018) is promising but requires further observation data. Pedestrian modeling was deemed outside the scope of this study, therefore pedestrian behavior has been modeled in a simplified way.
The implementation of a pedestrian model in the proposed framework is based on the scenario definitions that prescribe fixed pedestrian trajectories which are continuous in their velocities and accelerations, as further explained in Sections 3.3.2 and 3.5.2.

Scenario generation
Scenarios have been modeled based on target trajectories. A trajectory consists of path information and the corresponding time-dependent participant position. The simulated trajectory is then affected by the modeled physics within the simulation, i.e., road conditions and/or active safety system interventions.
As indicated in Section 1 conflict situations described by Lindman et al. (2011), indicate the road layout of the accident site and the intended lanes used by the participants. In this study, the OpenDRIVE road layout description has been used to derive paths, which are in accordance with described conflict situations. This represents the first part of this section. Further, it describes a range of time-dependent dynamics along the path points that can be assigned by motion states to derive target trajectories which can be simulated with the introduced simulation platform and models in Section 3.1.

Path derivation from OpenDRIVE road layouts
OpenDRIVE (Dupuis et al., 2010) is an open, XML-based format, in which road elements outline the main building blocks. Road elements are unique (defined by an ID) and contain information defining their physical characteristics. One property is the reference line (straight line, spiral, arc, etc.) upon which lanes, elevation profiles, road markings, and traffic signs are based. Each reference line R i is depicted through the coordinates of the origin (x i , y i ), the orientation α i , and its length l i with respect to the global coordinate system and additional geometrical information. Exemplary reference lines for two consecutive road elements are shown in Fig. 5. Road elements might be linked directly or through additional junctions, in the case of multiple roads. The graph-like network structure allows recursively traversing the network to find a route between a start and a destination road (denoted by their IDs).
Travel lanes are defined by boundaries that restrict the target trajectories of vehicle motion controllers. Lane borders are defined by their lateral profile and the road element's reference line. This geometrical description has been used to derive discrete inner B i and outer   Motion states are used to assign time information to each path point and to derive potential trajectories for the traffic participants.

Motion states
To define time history information for each agent's target trajectories, a kinematic-based approach has been used. From a given velocity v i and the acceleration a i in this point p i , the time increment Δt top perform the transition to p j can be calculated using basic kinematic relations. To state changing accelerations, a sequence of motion states has been defined. Motion states contain information about the accelerations within certain boundary conditions for braking and acceleration limits, for example. The initial motion state m 0 is described through a distance offset s 0 to the path point p 0 , the initial velocity v 0 , and the initial acceleration a 0 . Further, the motion states m i require the definition of an acceleration a i and the motion state's boundary condition. Using kinematically valid motion states, the timed transitions Δt of all consecutive points along the path P can easily be parameterized, which is a prerequisite for the automated generation of the scenario catalog.
As already mentioned in Section 3.1, target trajectories are based on the PCM definitions and require a fixed timestep Δt of 10 ms between consecutive target points p i and p j . The derived time Δt i for the transition between two path points does not follow this convention. Hence, it was necessary to linearly interpolate the derived path points P i and times t i with a constant Δt of 10 ms. The recalculated target points P are exemplified in Fig. 5.
A further issue that had to be taken into account is the restriction of reasonable motion states along a path. The PID based trajectory follower of openPASS incorporates a look-a-head time, which considers the path curvature and adjusts the brake throttle to lower the velocities accordingly. With the standard setting of the look-a-head time of 200 ms, problems were encountered when target velocities exceeded the limits due centripetal acceleration in curves. Finding a reasonable good look-ahead time is difficult and high values might lead to further side effects, such as steering too early. In order to cope with this issue, motion states are automatically recalculated if the target velocity exceeds the maximum permissible. Therefore it is necessary that the braking process (for vehicle navigation) is initiated early enough, to ensure that the target velocity never exceeds the physically feasible value. The time to activate the braking is determined by the theoretical braking distance (depending on the friction) to reach the threshold velocity.

Scenario evaluation
Output information such as agent positions (x, y) in the global coordinate system, velocities v → , as well as yaw angles ψ can be further used to manually visualize the outcomes of a simulation, observe possible malfunctions of the system or to investigate the reasons for a collision in the scenario. Collisions are determined by the geometrical overlap of the agent contours, which has been calculated for each time step in the simulation platform. Fig. 6 shows the position of both the car and the pedestrian at the collision time t 0 . Using the geometrical properties of the car and the pedestrian, the relative impact location to the center of the respective contour edge can be determined. The relative impact velocity v → r at t 0 is calculated using Equation (1).

Derived scenario catalog
The following section explains how the scenario catalog for the current study was generated. In the future, the catalog can be modified, extended, or derived results can be weighted with accident statistics.

Generic road network
The set of conflict categories was defined by the investigations of Lindman et al. (2011). In this study, we considered three main categories of conflict situations: car straight on (SCPPxxx), car left-turning (LTxxxx), and right-turning (RTxxxx). Each of these categories can further be split into sub-situations, with different pedestrian movement patterns. An example would be the abbreviation SCPPLSD, which stands for Straight Crossing Path, Pedestrian from (SCPP) Left (L) initially from Same Direction (SD). Considered conflict situations as well as their abbreviations are shown in Fig. 7. Statistics on German and British accident data  confirm the importance of these conflict situations since they are most common for fatal and severe pedestrian accidents. Conflict situations in which the pedestrian moves in the same or opposite direction of the cars driving on the road have been excluded in this study.
To model the car-pedestrian conflict situations requires a road network allowing cars to make three different movements (driving straight ahead and turning left and right). In addition, the road network has to be designed such that the pedestrian movements can be derived from it as well. Therefore, a pedestrian must be able to walk along the road and cross it, possibly at pedestrian crossings. Although pedestrian crossings can be modeled in OpenDRIVE, automated routing possibilities as described in Section 3.3.1 are limited. To cope with this issue, the developed road layout had to be extended with an additional pedestrian motion network to allow automated pedestrian routing. The created road network in this study is based on the test track of the European New Car Assessment Programme (Euro NCAP) vulnerable road user (VRU) AEB test protocol Euro NCAP (2019). It is symmetrical and its dimensions are 140 × 140 m. The geometries at the intersection and the width of the roadway are shown in Fig. 7. The geometrical properties of the pedestrian motion network are in line with the pedestrian's initial positions in the Euro NCAP test protocol and the corresponding sidewalks of the car road layouts. The implementation of a separate pedestrian motion network allows for other possible pedestrian motions not presented in the current study. Interpolated target trajectory points P defined by motion states (green). The illustrated acceleration of motion state 2 increases the distances between the target trajectory points. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.) Fig. 7 shows all conflict situations for the developed road layout and their respective abbreviations, which are further used to describe the simulated results in Section 4.

Agents' motion states
The scenario catalog combines the initial car and pedestrian velocities (v init ) with the different paths of conflict situations.
For the demonstration scenarios in the current study, pedestrian behavior was simplified. Therefore, only initial velocities as well as an offset time t start were used to describe the (baseline) motion states of the pedestrian. The range for pedestrian velocities includes pedestrian walking and running paces and these have been chosen following observations on crossings (Almodfer et al., 2017).
The generic road layout limits the velocity for the car according to feasible lateral acceleration as described in Section 3.3.2 and the prescribed speed limit. Braking at a specific velocity in a curve must be feasible from the initial position ane limit deceleration of 7 m/s 2 to achieve the target velocity. Besides having to brake due to the curve as described in Section 3.3.2, braking due to driver interaction with the surroundings was not considered.
Combinations of the individual parameters are possible according to their ranges. The ranges of these parameter can be found in Table 1.

System parameters
The pedestrian definition in the developed framework reflects the dimensions of the pedestrian dummy in the Euro NCAP test protocol (Euro NCAP, 2019) and has a width of 0.54 m and a length of 0.28 m.
The car model on which the study was conducted has a length of 4.9 m, a width of 1.9 m, and the weight is 1750 kg. The center of gravity is located 1.42 m behind the front axle at a height of 0.78 m and 2.45 m behind the leading edge. Its wheelbase is 2.84 m and the track width is 1.63 m.
For the AEB equipped car, exemplary AEB sensor parameters have been selected in the study described in Section 4 as the following:

Table 1
The virtual testing catalog was determined by a variation of conflict situation, the initial velocities v init of the pedestrian and the car as well as a timed offset for the pedestrian t start . The combination of these parameters resulted in a total number of 46,080 different virtual testing scenarios for the 16 considered conflict situations.  Gruber et al. (2019) and Wimmer et al. (2019).

Results
In order to determine the influence of the AEB on the derived scenarios, the catalog was initially simulated with a car model with or without an AEB or driver reaction functions (baseline cases). The baseline cases serve as a reference for the simulation of the same virtual scenarios with an AEB system (AEB cases) as shown in Figs. 2 and 3. Car configurations, such as control parameters or geometrical properties, have not been changed between the baseline and the AEB case. The results obtained from the demonstration efficacy assessment are summarized in this section, and contain the shift in accident configurations (impact location and relative velocities).
Out of 46,080 baseline simulations, 6403 ended in a collision. This number was reduced to 5270 for the AEB cases. For the investigation of collision velocities and impact points, the cases were divided into subsamples based on the impact location. Of particular interest are scenarios in which the pedestrian was hit by the car front, identified as baseline/AEB front cases. A total of 4075 accidents occurred in the baseline front case compared to 3091 for the AEB front case, representing a reduction of 24.1% for this type of accident. A closer look at the distribution among the conflict situations show that the conflicts are not equally distributed among the collision types. Details of the observed data are given in Table 2.
In the context of the individual conflict situations, it can be stated that the AEB is more effective in situations where the car is traveling in a straight path (SCPPL, SCPPLOD, SCPPLSD, SCPPR, SCPPROD, SCPRSD). In turning situations, the implemented AEB system showed hardly any effects in terms of crash avoidance or mitigation. In some turning conflict scenarios the situation worsened in terms of increased collisions or lack of severity reduction. The mean collision velocity v coll was reduced from 9 m/s to 7.8 m/s resulting in an improvement of 13.5% for front cases. The distribution of the impact location was not significantly affected by the AEB system. Both distributions can be seen in Fig. 8. Fig. 9 shows the proportion of non-conflicting scenarios and the front baseline cases in which the AEB has either avoided, reduced, or not had an effect on the collision velocity. Detailed statistics for these subsets can be found in Table 3.
Further case-based investigations of frontal crash scenarios reveal that the AEB does not address all the baseline front cases. This means that under certain circumstances, the intervention of the AEB led to a collision. In Fig. 10 a direct comparison between a simulated baseline and the AEB scenario is shown. For this scenario, the triggered emergency brake led to a conflict with the car front, which was not observed in the baseline case.

Discussion
Most previous studies have applied real-world accident scenarios to assess the efficacy of AEB systems (Lindman, Ö dblom et al., 2010;Rosén et al., 2010;Hummel et al., 2011;Jeppsson et al., 2018;Gruber et al., 2019). In contrast, this study uses the road network description format OpenDRIVE to define a scenario catalog based on possible car-pedestrian conflict situations. The discussion includes an examination of the advantages and limitations of the methods used to generate and simulate scenarios in Section 5.1 and an outlook to future applications and extensions in Section 5.2.

Methodology aspects
Sophisticated and open formats to describe road networks, such as OpenDRIVE, allow modeling more complex road geometries than the implemented Euro NCAP road configuration shown in Fig. 7. A complete implementation of such a standard naturally requires additional effort and was outside the scope for this investigation. The developed Table 2 Overall and front conflict cases split into the different conflict situations. The evaluation shows that the AEB is comparatively effective for situations in which the car is traveling straight ahead since the majority of collisions could be avoided. For some conflict situations, the number of accidents increased. methodology, as described in Section 3.3.1 is, however, capable of extracting driving paths from common generic road layouts, including road intersections. An investigation into the safety aspects of more complex road layouts should be considered in the future.
To the best knowledge of the authors, this study is the first published application that makes use of the simulation platform openPASS and primarily benefits from its main architectural design decision to be open, transparent, and fast in its execution. In comparison with other frameworks, the assessment of active safety systems is a central part of the entire software design. This effect was shown within the study, in which almost 100,000 scenarios have been simulated within a day on conventional hardware. Furthermore, their execution can easily be parallelized with other computing units in the future. The framework is being constantly modified and the freely accessible models are still relatively simple. One example can be seen in the modification of the motion states described in Section 3.3.2, to limit the car velocity to suitable lateral acceleration levels. This addresses the shortcoming the original trajectory follower (PID controller) cannot adequately compensate excessive accelerations and decelerations.
Single, defined, accident scenarios from accident reconstructions are often used to investigate the sensitivity of certain active safety system parameters (Hummel et al., 2011;Hamdane et al., 2015;Gruber et al., 2019). As indicted by Menzel et al. (2018), it is however necessary to vary "critical" parameters within the context of the original conflict, to explore possible limits of the boundary conditions of an active safety system. Accident reconstructions contain some inaccuracies and discretely defined scenarios by themselves may not be complete enough to address the stochastic patterns of the accident participants and the environment.
The results in Section 4 showed that some scenarios, which did not lead to a conflict in the baseline, resulted in a conflict once the AEB system had been introduced. In this study, this might be possible due to the shortcomings of the introduced AEB system. Motion predictions were implemented in a simplified manner, extrapolating the current velocity and direction to the future without taking further temporal information about the past into account. For the enhancement of the AEB system mode, it would be helpful to extend it with appropriate predictive control strategies as shown in Kooij et al. (2014). Further, the sensor of the simulated generic AEB system has been modeled using a somewhat simplified raytracing algorithm, which takes neither weather effects nor latencies for data acquisition and pedestrian detection into account. Delays for data processing and incorporating miss rates for pedestrian detection, which is a common performance measure, would further enhance the sensor model. The influence of sensor system delays and miss rates would result in reduced efficacy predictions, as concluded in Hamdane et al. (2015).
The purpose of this framework is for the user to evaluate their specific systems. Sophisticated manufacturer-specific system models can easily be incorporated and tested on the generated scenario catalogs. As shown in this paper, sensitivity studies incorporating slightly changing scenarios lead to a more holistic assessment of a particular conflict, and can be used to derive further requirements on the active safety systems as well as explore its limits.
In the context of the generated scenario catalog, driving paths are currently assumed to be lane centered. Further, no additional braking, from a model of a driver reacting to an impending crash has been taken into account, nor an additional pedestrian reaction, i.e., changing speed or direction. Configurations in which driver and pedestrian fail to notice each other might be true for a significant subset of accidents (Soni et al., 2013;Schachner et al., 2020). However, these distributions would only be revealed by a further statistical investigation of real-world accidents. In addition to the velocities, the relative position of the pedestrian to the car was varied by changing the initial starting time t start . The relevance of this parameter on AEB efficacy has previously been highlighted by Hamdane et al. (2015).
The simulated results in Section 4 showed a comparable outcome to previous investigations (Lindman, Ö dblom et al., 2010;Rosén et al., 2010;Hummel et al., 2011) (reduction by 24.1%). Nevertheless, the results should be understood as a demonstration of the framework, since a number of assumptions regarding scenario frequency and participant behavior have been made without taking actual accident statistics into account. The described conflict reduction by 24.1% for the front cases in Section 4 is based on the assumption that the occurrence of each scenario is equally distributed, the same applies to the speed reduction of 13.5%. This assumption also applies to the comparison within a conflict situation with other studies such as shown by Kovaceva et al. (2020). The mean has been reduced from 9 m/s to 7.8 m/s; (b) normalized impact location to the front center of the car. The impact location did not change significantly. Fig. 9. Out of the 46,080 scenarios, 4075 ended in a collision with the car front (8.8%). The AEB avoided 2.7% of these conflicts and 5.3% were reduced in severity. In 0.8% of the cases, the AEB did not show any mitigating behavior.
The avoidance of conflict situations in which the car is traveling straight ahead and the pedestrian is crossing the road was greatest in this study. On the other hand, avoidance of the class of conflict situations in which the car is turning was significantly lower. With regard to turning conflict situations, the mean collision velocity in the comparative study was lower, which might be an indicator of potentially lower collision velocities in general.

Future work
The main advantage of deriving scenarios based on conflict situations is that developed methodology can be applied transparently and holistically, in combination with real-world accidents, without the need to reveal entire accident records. Dedicated accident statistics can thus be used to weight scenarios according to the occurrence probabilities of individual parameters (such as initial velocities and conflict situations) to derive a catalog of virtual testing scenarios or to weight respective results, and conclude impact locations and velocities of remaining accident scenarios. Weighted results would further allow concluding the remaining injury risk from the simulated impact locations as done by Lindman et al. (2011 and Jeppsson et al. (2018).
The motion paths of different agents (cars and pedestrians) are described such as to allow the implementation of different characteristics (acceleration, trajectory changes, etc.). As already indicated, individual pedestrian movement is rather complex and influenced by multiple factors. In this paper, the sample road layout defines the participants' paths and has therefore a significant influence on the simulated results. The pedestrian road network was designed to meet the layout of the Euro NCAP test protocol and foresee pedestrian sidewalks and crosswalks. Nevertheless, in real-world scenarios, the entire crosswalk is used by pedestrians (Fugger et al., 2001) and in certain circumstances, they also move on roads that can be foreseen for other road participants (Yannis et al., 2013). The sample road layout defines the participants' paths and has therefore a significant influence on the simulated results. It would be beneficial to model pedestrian behavior stochastically (Huang et al., 2018) including start and goal regions defined on the road layouts as well as to use the agent-based simulation properties of openPASS to model pedestrian pre-crash behavior as observed in Soni et al. (2013) and Schachner et al. (2020). Further, accident database investigations show that sight obstructions, such as parked cars or buildings, represent a common cause for pedestrian accidents  foreseen in the PCM like scenario definitions as well as in OpenDRIVE. Therefore, incorporating sight obstructions would be of benefit in future studies, since they are assumed to be most challenging for AEB systems to detect .
Transparency plays an increasingly important role in the investigation of future safety performance, hence standards for comparability between studies are becoming increasingly more important. Open and freely accessible simulation platforms like openPASS attempt to safeguard transparency. Corresponding studies, which compare the simulated outcomes (Wimmer et al., 2019) of openPASS and other state-of-the-art simulation platforms (Wille and Zatloukal, 2012;Kolk et al., 2016), would increase the acceptance of openPASS. The same applies to the scenario descriptions, which guarantee transparency and exchangeability through the development of formats such as Open-Scenario (VIRES Simulationstechnologie GmbH, 2015).

Conclusions
This work introduces an open-source approach that uses the road network description format OpenDRIVE to develop a scenario catalog. The drawbacks of the application of case-based real-world accidents have been identified and addressed. A catalog in excess of 46,000 scenarios was derived with the developed methodology for a single pedestrian crossing, which has been rapidly simulated and evaluated. Real-world accident data can be incorporated in the scenario catalog. It Table 3 Evaluation of the 4075 baseline front cases split up into the different conflict situations and into the disjunct subsets avoided by AEB, reduced collision velocity by AEB, and no AEB intervention. The evaluation shows a significant conflict reduction for situations in which the car is traveling straight ahead. is however not limited to the reconstructed cases only. Also scenarios in between observed real-world cases should be considered, which can be done with the introduced method. The introduced open-source tool allows users to combine generic and real-world scenarios to assess conflict situations in a manifold and systematic way. The catalog can be used to reveal malfunctioning and possible shortcomings of active safety measures and for the determination of remaining future pedestrian accidents and their configurations, which should be addressed by passive safety measures.

Conflict of interest
None declared. Fig. 10. Direct comparison between the baseline and AEB simulation. For the AEB case, the brake was triggered, which led to an accident in comparison to the baseline case. The pedestrian's initial velocity v init is 2 km/h, its timed offset 3.1 s and the car's initial velocity was set to 28 km/h.