Accident Analysis and Prevention 214 (2025) 107962

Contents lists available at ScienceDirect

Accident Analysis and Prevention

journal homepage: www.elsevier.com/locate/aap

Generation of critical pedestrian scenarios for autonomous vehicle testing

Lan Yang a , Songyan Liu a, Shuo Feng b,*, Hong Wang c , Xiangmo Zhao a , Guangyue Qu a ,
Shan Fang a
a
School of Information Engineering, Chang’an University, Xi’an 710064, China
b
Department of Automation, Tsinghua University, Beijing 100084, China
c
Intelligent Vehicle Design and Safety Research Institute, Tsinghua University, Beijing 100084, China

A R T I C L E I N F O A B S T R A C T

Keywords: Current autonomous vehicle (AV) testing scenarios predominantly focus on interactions between AV and sur­
AV testing rounding vehicles, with limited consideration given to high-risk pedestrian interactions. This paper presents a
Pedestrian testing scenario method for generating critical test scenarios specifically designed for pedestrian-oriented evaluations. First,
Vehicle-pedestrian interaction
microscopic traffic data were collected from 12 signalised intersections in 4 cities across China. By extracting
Scenario generation
overlapping vehicle and pedestrian trajectory data within the same spatiotemporal context, a vehicle–pedestrian
interaction scenario library was created. Second, a three-stage autonomous emergency braking model was used
to simulate the decision-making and control processes of AV, replacing the vehicle agency in the original scenario
library. In addition, the artificial potential field method was applied to assess real-time interaction risks, enabling
the identification of high-risk scenarios. A pedestrian-oriented critical test scenario generation framework was
then developed, defining key decision variables such as speed differences, relative lateral distances, and relative
longitudinal distances between pedestrians and vehicles. An importance sampling function, incorporating both
scenario exposure frequency and interaction risk, was designed to generate critical scenarios. The process was
further refined with an auxiliary objective function to guide the search direction. To improve computational
efficiency, swarm optimisation and flood-fill algorithms were employed. Using this method, 50 high-value
vehicle–pedestrian interaction test scenarios, characterised by high exposure frequency and risk, were gener­
ated. These scenarios encompass diverse and high-risk interaction dynamics, providing robust support for high-
fidelity pedestrian safety testing of AV.

1. Introduction pedestrians accounting for 21 % of these fatalities, making pedestrians

highly vulnerable road users (VRUs) at an elevated risk of traffic acci­
In recent years, AV have demonstrated significant technical advan­ dents. As a result, accurate pedestrian identification, along with
tages in enhancing traffic safety, alleviating congestion, reducing envi­ appropriate decision-making and control, is crucial for the widespread
ronmental pollution, and decreasing energy consumption, positioning deployment of AV. Pedestrians display unique movement characteristics
them as an inevitable trend in the evolution of the automotive industry. and complex, variable behaviour patterns, particularly when crossing
Governments and companies worldwide are intensifying their research streets and intersections (Sengupta et al., 2024), which introduces sig­
and development efforts in AV technology, recognising it as a key op­ nificant uncertainty into AV decision-making and control processes
portunity for industry transformation and upgrading, marked by a shift (Wang et al., 2023). Consequently, extensive AV testing focused on
from large-scale production to high-quality growth. pedestrian interactions is essential (Rezwana and Lownes, 2024; Rasouli
According to SAE’s automation levels, Level 3 and above (L3+) AV and Tsotsos, 2020).
do not require drivers to consistently perform driving tasks, potentially Before being deployed on public roads, AV must undergo a series of
reducing risks associated with human error and enhancing pedestrian rigorous testing procedures to ensure their reliability and safety
safety. Meanwhile, the World Health Organization’s Global Status (Sohrabi et al., 2021). The tests primarily focus on the response of AV in
Report on Road Safety (Global status report on road safety, 2023) various complex traffic, which poses more stringent requirements on
highlights that global road fatalities reach 1.35 million annually, with testing methods, and has become a key challenge in the field of AV

* Corresponding author.
E-mail address: [email protected] (S. Feng).

https://doi.org/10.1016/j.aap.2025.107962
Received 1 December 2024; Received in revised form 10 February 2025; Accepted 10 February 2025
Available online 18 February 2025
0001-4575/© 2025 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
L. Yang et al. Accident Analysis and Prevention 214 (2025) 107962

testing. International standards related to AV testing, such as SAE J3016, model was employed to simulate the decision-making and control
ISO 21448 (International Organization for Standardization (ISO), 2022), processes of AV during interactions with pedestrians, replacing
ISO 26262 (International Organization for Standardization (ISO), 2018), the vehicle agency in the original scenario library. Additionally,
ISO 34501 (International Organization for Standardization (ISO), 2022), the artificial potential field (APF) method was used to assess real-
and ISO 34502 (International Organization for Standardization (ISO), time risk fields for both vehicles and pedestrians. This approach
2022), have been proposed consecutively, aiming to address and explore enables more precise identification of hazardous boundary sce­
the standardized definitions and testing content for functional safety, narios during vehicle–pedestrian interactions, thereby improving
safety of the intended functionality (SOTIF) and testing scenarios. the reliability of risk assessments in critical situations.
Among them, ISO 34502 introduces a multi-pillar approach-based AV
testing system, which encompasses virtual testing, closed-field testing, 2. Literature review
and open-road testing. The three testing tools and scenarios complement
each other to achieve the goal of comprehensive testing coverage. Vir­ In recent years, researchers globally have been exploring viable
tual testing is the foundational component of the AV test system, aiming methods for testing the safety of AV. Extensive safety testing of AV in the
to maximize testing scenarios that are difficult or impossible to replicate early stages primarily focused on continuous acceleration tests in
in controlled closed-field or open-road environments. Specifically, a various vehicle interaction scenarios such as merging, exiting, car-
wide range of high-risk, complex, critical, and even corner cases must be following and lane-changing, aimed at identifying the safety bound­
addressed during the virtual testing phase to evaluate the upper limits of aries of AV (Zhao et al., 2018; (Wei et al., 2024). However, once the
AV safety performance effectively. basic tests are satisfied, the scope should be expanded to include safety
Most of AV testing research focuses on scenarios involving in­ testing for other traffic participants (TPs) in real traffic scenarios. In
teractions between the tested AV and surrounding vehicles, encom­ recent years, attention has increasingly turned to pedestrian testing for
passing typical, hazardous, and accident scenarios (Baby et al., 2024; Li AV. Existing research on AV pedestrian testing includes pedestrian
et al., 2024; Zhou et al., 2023). However, it is essential to emphasise that scenario testing based on standard regulations, pedestrian scenario
pedestrians are among the most unpredictable and critical dynamic testing based on simulation tools, and methods for generating pedestrian
factors that vehicles must account for when operating in real-world testing scenarios for AV.
traffic environments. This is particularly evident at urban intersections Before mass production, AV need to undergo pedestrian collision
and high-traffic areas, where pedestrian flow is dense and behaviours avoidance tests based on specific standards and regulations, such as
are complex and varied. During peak hours, frequent crossings and Europe’s ECE R127 and China’s GB/T 38186–2019 (UNECE and GB/T
unpredictable mid-street movements further complicate real-time deci­ 44719). Such tests are generally conducted in closed test fields by setting
sion-making for AV, escalating both the difficulty and risk. In such standardized scenarios and using dummies to establish initial positions,
scenarios, comprehensive pedestrian testing allows AV to adapt to speeds, and collision angles with vehicles. However, the limited condi­
diverse pedestrian behaviours and high-density environments, effec­ tions of closed test fields make it challenging to fully replicate complex
tively reducing pedestrian-related accidents and enhancing the system’s and random pedestrian scenarios (Zhang et al., 2023).
ability to respond to complex traffic situations. Consequently, this type Consequently, some researchers have been investigating the preset­
of testing is vital for ensuring the safe deployment of AV in densely ting of pedestrian behaviours (such as walking speed and path) in
populated urban areas. simulation conditions to create more complex testing scenarios, assess­
To address the gap in AV testing research, this paper introduces a ing the response capabilities of AV (China New Car Assessment Pro­
framework for generating critical pedestrian test scenarios, leveraging gramme (C-NCAP), 2024) (European New Car Assessment Programme
microscopic traffic datasets from signalised intersections in large and (Euro NCAP), 2023). Studies have proposed using VR simulation to build
medium-sized cities across China. The framework is designed to virtual environments that can simulate complex pedestrian behaviours
generate high-value scenarios that involve frequent and high-risk in­ and reactions (Tran and Parker, 2024) (Kwon et al., 2022), allowing AV
teractions between vehicles and pedestrians, providing valuable insights to conduct extensive pedestrian tests in safe and controllable environ­
for high-fidelity AV testing. ments. Software-in-the-Loop (SIL) testing verifies the performance of the
decision-making module of autonomous systems by simulating pedes­
(1) This study presents a framework for generating critical pedestrian trian scenarios in a virtual environment, particularly when pedestrians
testing scenarios for AV. The relative speed and distance between approach or interact (Yu et al., 2023; Angulo et al., 2023). Model-in-the-
vehicles and pedestrians are defined as the decision variables for Loop (MIL) testing simulates the control logic of autonomous systems
the scenarios. An importance sampling function, based on sce­ through computer models, which is particularly suitable for testing
nario exposure frequency and interaction risk levels, is devel­ pedestrian recognition and path planning models (Bruggner et al.,
oped, with an auxiliary objective function to guide the 2021). Hardware-in-the-Loop (HIL) testing connects vehicle sensors and
identification of critical scenarios. This framework effectively control systems to a simulation platform to assess system responses
generates a diverse range of high-risk vehicle–pedestrian inter­ when pedestrians approach or interact (Zhu et al., 2024).
action scenarios in urban intersection environments. Scenario-driven testing methods for AV have become an industry-
(2) To examine the significant impact of pedestrian characteristics on recognized solution for addressing the performance testing challenges
AV in China, this study collected microscopic traffic datasets from of AV. Such methods aim to conduct targeted tests on high-value sce­
12 signalised intersections across 4 large and medium-sized cit­ narios with minimal testing efforts. Lin et al. (Lin et al., 2025) and Fei
ies: Xi’an, Tianjin, Chongqing, and Changchun. The datasets et al. (Fei et al., 2024) emphasise the critical role of control engineering
encompass key elements for AV test scenarios, including vehicles, and big data-driven vehicle perception technologies in intelligent con­
traffic signals, and various VRUs. By filtering vehicle trajectory nected vehicles, providing significant technical support for the safety
data influenced by traffic signals and incorporating trajectories of evaluation of autonomous systems and pedestrian testing scenarios.
normal behaviour, red-light violations, accelerated crossings, Similar to vehicle-to-vehicle interaction tests, pedestrian testing sce­
evasive manoeuvres, and actions around vehicles, an original narios for AV seek to extract as many rich and realistic high-interaction
scenario library was established. This library reflects the unique vehicle–pedestrian scenarios as possible from a vast array of real traffic
characteristics of Chinese drivers and pedestrians, specifically situations. Current methods for generating pedestrian scenarios in AV
tailored for pedestrian testing of AV. primarily involve constructing scenarios from traffic accident data and
(3) To accurately identify high-risk vehicle–pedestrian interaction generating scenarios based on natural driving data (NDD). Researchers
scenarios, a three-stage autonomous emergency braking (AEB) like Lenard et al. (Lenard et al., 2018) have used clustering methods to

2
L. Yang et al. Accident Analysis and Prevention 214 (2025) 107962

Fig. 1. The overall framework of the research.

identify 6 typical logical scenarios representing 86 % of the cases from generation. As the application of large language models continues to
9,360 car-pedestrian collision incidents. Researchers such as Su et al. (Su rise, Kuang et al. (Kuang et al., 2024) and Qu et al. (Qu et al., 2023)
et al., 2017) and Liu et al. (Liu et al., 2012) have utilized systematic explore their potential in traffic scenario generation and decision-
clustering methods based on NDD to identify 4 and 5 classes of repre­ making support, offering new technological pathways and support for
sentative pedestrian traffic conflict logical scenarios, respectively. enhancing the performance of autonomous driving systems and their
Similarly, researchers have applied clustering analysis algorithms to safety assessment in complex traffic environments. Researchers like
extract typical logical scenarios from hundreds of car-to-bicycle collision Spooner et al. (Spooner et al., 2021) have proposed GAN-based methods
incidents (Hu et al., 2018; Cao et al., 2020; Sui et al., 2017). Although for generating pedestrian crossing scenarios, enabling automated
the scenario construction methods are based on real accident cases, the perception and decision systems to more effectively identify and
completeness of scenario reproduction is limited by the lack of data prior respond to pedestrian crossings. Furthermore, methods for scenario
to the accidents. Researchers like Ma et al. and Xu et al. (Xu et al., 2021; enhancement simulation have been proposed, which generate critical
Ma et al., 2021) have focused on constructing and validating pedestrian- scenarios or edge cases to test the performance of automated systems in
vehicle and vehicle-to-vehicle interaction scenarios based on NDD. The high-risk or low-probability situations (Klück et al., 2019; Sun et al.,
adoption of real NDD to create realistic pedestrian behaviour models 2023; Zhou et al., 2024; Muktadir and Whitehead, 2024).
and interaction scenarios has been proposed (Zhu et al., 2021; Tian The generation of high-value pedestrian testing scenarios must
et al., 2023). Currently, popular generative models include Generative ensure comprehensive coverage while addressing uncertainties in traffic
Adversarial Networks (GANs) and Variational Autoencoders (VAEs), situations and pedestrian behaviours, such as the randomness of
which model distributions and generate new data from large datasets pedestrian actions and variations in traffic signals. While pedestrian
(Krajewski et al., 2018; Zhang et al., 2021; Abeysirigoonawardena et al., scenario testing for AV has gained increasing attention, research on
2021; Mathieu et al., 2019; Yang et al., 2023). Li et al. (Li et al., 2025) critical pedestrian testing scenarios remains limited. This paper pro­
conducted a systematic review on the application of GANs in traffic state poses a framework for generating AV pedestrian testing scenarios, with a
prediction, highlighting their potential in enhancing AV scenario focus on extracting vehicle–pedestrian interaction characteristics from

3
L. Yang et al. Accident Analysis and Prevention 214 (2025) 107962

Fig. 2. Selected signalized intersection of 4 cities in China.

typical traffic situations in China. constitute the critical pedestrian testing scenarios library.

3. Research methodology 4. SIND dataset

In this paper, we develop a pedestrian-oriented critical testing sce­ 4.1. Data recording and statistics
nario generation framework based on the principle of importance sam­
pling and the APF method (Wang et al., 2016). The accuracy and The behaviour of vehicles and pedestrians at signalized intersections
effectiveness of the critical scenarios generated are validated through is regulated by periodic traffic signals to minimize conflicts. However,
comparative experiments. Fig. 1 illustrates the overall framework pro­ numerous studies have shown that traffic accidents at intersections have
posed in this study. persisted due to violations of traffic rules by both vehicles and pedes­
Microscopic traffic datasets are collected from 12 signalized in­ trians (Shaon et al., 2018). Densely populated large and medium-sized
tersections across 4 major cities in China, including different kinds of cities in China exhibit complex pedestrian behaviours at signalized in­
pedestrian crossing behaviours which is used to form the original sce­ tersections, characterised by frequent interactions between pedestrians
nario library. Decision variables are subsequently selected as feature and vehicles. Such scenarios present considerable challenges for the
descriptors to characterize vehicle–pedestrian interaction scenarios. deployment of AV in China. To address this, it is crucial to conduct
Inspired by the novel framework proposed by Feng et al. (Feng et al., statistical analyses, surveys, and research on pedestrians, vehicles, and
2021; Feng et al., 2021), which integrates exposure frequency and their interactive behaviours at real intersections for AV testing.
manoeuvre challenges of scenarios through importance sampling for the Currently, there is no open-source, high-quality microscopic traffic
vehicle-vehicle interaction scenarios generation, we extend this princi­ dataset available from signalized intersections in China to support
ple to the pedestrian testing scenarios generation in this study. The research on AV test scenarios. In this study, trajectory data of SIND
definitions of exposure frequency and risk level in this paper are as (Signalized Intersection Dataset) were independently collected from 12
follows: signalized intersections across 4 cities in China, providing real-world
behavioural data of VRUs to support the analysis of dynamic features
• Exposure frequency: the occurrence probability of scenario in real in complex vehicle-to-everything interactions.
traffic environments. Data were collected from 12 signalized intersections across 4 cities in
• Vehicle-pedestrian interaction risk: the collision probability between China—Xi’an, Tianjin, Chongqing, and Changchun. For the purposes of
the vehicle and pedestrian. this study, four intersections were selected to generate critical pedes­
trian testing scenarios, as depicted in Fig. 2. Specifically, intersection 1 is
The exposure frequency is determined by analysing the statistical located in Xi’an, intersection 2 in Changchun, intersection 3 in
distribution of decision variables within the original scenario library. Chongqing, and intersection 4 in Tianjin. The data were recorded using a
Scenario risk is initially assessed using a three-stage AEB model as a DJI Mavic 2 drone equipped with a 4 K (3840 × 2160) camera, capturing
surrogate model (SM), which replaces vehicle actions in the original footage at a frequency of 29.97 Hz. The drone hovers at a high altitude,
scenario library with AV responses to pedestrian crossings. Following minimizing its impact on VRUs, while its high-definition camera cap­
this, an APF-based risk model is employed to evaluate the risk levels of tures natural behaviours (Zhan et al., 2019). In addition, mobile phones
vehicle–pedestrian interaction scenarios with the integrated SM. After were placed at two opposing corners of each intersection to record
defining the metrics for critical scenario extraction, an auxiliary func­ traffic light statuses and capture pedestrian crossing details more
tion and an optimisation search algorithm are developed to enhance the clearly. All intersections are controlled by two-phase signals, with left-
efficiency of generating critical scenarios from the original scenario li­ turning and straight-moving vehicles sharing the same green light
brary. Ultimately, the scenarios generated using the proposed method phase. This setup often leads to frequent interactions and conflicts

4
L. Yang et al. Accident Analysis and Prevention 214 (2025) 107962

Table 1
Details of the selected signalized intersections.
No. City Number of pedestrians Number of bicycles Number of motorcycles Number of Number of Number of Number of
tricycles cars trucks buses

1 Xi’an 4447 396 3862 345 8090 199 102

2 Changchun 1267 237 1339 99 12,074 622 408
3 Chongqing 1157 3 439 27 2171 46 5
4 Tianjin 1652 2807 3360 479 4792 107 51

Fig. 3. Typical conflicts of the vehicle–pedestrian interaction at signalized intersections.

between pedestrians and vehicles, facilitating the extraction of between vehicles and pedestrians. Vehicle-pedestrian conflict usually
numerous vehicle–pedestrian interaction scenarios. The dataset en­ occurs when the distance between pedestrians and vehicles decreases to
compasses a variety of motorized vehicles, including cars, buses, and a critical threshold, potentially leading to collisions or prompting
trucks, as well as VRUs such as pedestrians, cyclists, and motorcyclists. evasive actions by either party to prevent an accident (Yang et al.,
Data were collected during peak hours—morning, noon, and evening­ 2022).
—on both weekdays and holidays. Weather conditions, including sunny, Based on the SIND dataset, the following rules are adopted to extract
cloudy, and post-rainy days, were also considered in the data collection and form the original vehicle–pedestrian interaction scenario library:
process. This study constructed the SIND dataset through steps of
detection, correction, tracking, and post-processing (Xu et al., 2022). (1) A potential conflict area exists between pedestrian and vehicle.
The details of the selected sites are shown in Table 1. (2) From the onset of the interaction until one party exits the po­
In the data from Xi’an and Chongqing, pedestrians constitute the tential conflict area, the timestamps of the vehicle and the
largest proportion among VRUs, while in Changchun, the number of pedestrian are precisely synchronized.
pedestrians is comparable to that of motorcycles. In Tianjin, pedestrians (3) At the initial moment of interaction, neither the pedestrian nor
rank second only to two-wheelers. Additionally, the dataset includes the vehicle is within or has exited the conflict area.
detailed characteristics of pedestrian crossings, such as variations in
walking speed, trajectories, adherence to traffic regulations, observation Scenarios in which pedestrian and vehicle trajectories overlap within
and judgment of traffic signals, and group behaviours during road the same time frame are extracted as valid interaction scenarios from the
crossings. The diverse data and pedestrian portrait offer valuable sup­ original scenario library. Fig. 3 presents several common types of
port for the pedestrian-oriented testing and relative research. vehicle–pedestrian interaction conflicts at intersections. It highlights
key situations, such as when pedestrians cross paths with vehicles
making left or right turns, or when pedestrians and vehicles move in
4.2. Data processing for vehicle–pedestrian interaction scenarios
conflicting directions within the intersection. The original scenario li­
brary is primarily constructed by extracting the scenarios depicted in
To extract the dangerous scenarios of vehicle–pedestrian interactions
Fig. 3 from the collected datasets. Each conflict type is carefully
from original dataset, it is first necessary to identify conflict areas

5
L. Yang et al. Accident Analysis and Prevention 214 (2025) 107962

intersections, where pedestrians interact with approaching vehicles

from either the left or right. The dynamic interaction between pedes­
trians and vehicles, alongside other environmental factors, forms the
core of the pedestrian-vehicle interaction scenario. This interaction can
be divided into three phases: the initial encounter between pedestrians
and vehicles, the crossing phase where pedestrians move while vehicles
from either side continuously approach, and the final phase when one
party exits the conflict zone.
AV’s driving behaviour is critical when a pedestrian crossing ahead,
improper actions may lead to potential collision risks (Lanzaro et al.,
2023). The primary variables include the pedestrian and vehicle
movement characteristics, such as the crossing speed and coordinates of
both pedestrians and vehicles. Therefore, the vehicle’s longitudinal
speed vveh , the pedestrian’s crossing speed vped , their relative longitudi­
nal distance dlon , relative lateral distance dlat , and the speed difference
between the pedestrian and the vehicle diff v− p are selected as the deci­
sion variables representing the characteristics of the scenario, as shown
in Eq. (1). The decision variables in the pedestrian-vehicle interaction
scenario is shown in Fig. 4.
[ ]
x = vveh , vped , dlon , dlat , diff v− p (1)

5.2. Importance function

Fig. 4. The decision variables of the vehicle–pedestrian interaction scenario.

In the process of generating scenarios, a criticality assessment

identified and categorized, ensuring that all possible pedestrian-vehicle
method is required to determine the importance of each scenario and
interaction scenarios are represented. Additionally, special cases, such
filter out critical scenarios from the original scenario library. Currently,
as traffic violations by pedestrians or vehicles—though less fre­
most researches focus on the safety aspect, using indicators such as
quent—are also taken into account. Violations include cases in which
surrogate safety measures (SSMs). They do not consider the occurrence
pedestrians disregard traffic signals or vehicles fail to yield to pedes­
probability of scenarios which means these scenarios are prone to occur
trians. Although such occurrences are infrequent, their inclusion is
and typical in real traffic. According to Automotive Safety Integrity
essential for the development of a comprehensive scenario library for
Levels (ASIL) of the ISO 26262, three decisive factors, such as severity,
testing purposes. Ultimately, the extraction resulted in 555 valid sce­
exposure, and controllability, are defined to assess the risks associated
narios from Changchun, 805 from Chongqing, 846 from Tianjin, and
with automotive system components. Therefore this paper measures the
3,679 from Xi’an. These variations ensure a broad and diverse repre­
scenario criticality V(x|θ) with the exposure frequency and the risk level.
sentation of pedestrian-vehicle interaction scenarios.
The scenario exposure frequency P(x|θ) is obtained through the analysis
of the original vehicle–pedestrian interaction data, and the scenario risk
5. Critical pedestrian testing scenario generation
level P(S|x, θ) is assessed by the SM and a risk model with a certain
threshold. The importance function is defined as the product of two
5.1. Decision variables
factors, as shown in Eq. (2).

In the context of AV testing, a scenario refers to the interaction be­ V(x|θ) = P(S|x, θ)P(x|θ) (2)
tween the tested vehicle and environment, representing a sequence of
events over time (Cai et al., 2022). It defines the vehicle’s actions and 5.2.1. Exposure frequency
events as test objectives. A scenario includes both dynamic interactions The exposure frequency refers to the occurrence probability of sce­
among traffic participants and static traffic conditions (Ulbrich et al., nario which is determined from the proportion of scenarios within each
2015). Bagschik et al. (Bagschik et al., 2018) proposed a five-layer unit relative to the total. In the calculation, the speed difference diff v− p
framework to describe scenario elements, later expanded by Sauerbier between the pedestrian and vehicle at the initial encounter in all sce­
et al. (Sauerbier et al., 2019) to include a sixth layer with digital in­ narios is within the range (0,16). The range for lateral distance dlat is (0,
formation. Existed researches on scenario construction suggest that the 30], and the range for longitudinal distance dlon is (0, 40]. All parameter
elements of AV testing scenarios can generally be divided into two ranges are derived into discrete units from the original dataset with a
categories. The first category consists of fixed elements, which are step size of 1 for discretization.
determined by the Operational Design Domain (ODD), such as road
type, lane number, and other static factors. These elements define the 5.2.2. Risk model of vehicle–pedestrian interaction scenario
environment in which the AV is expected to operate and impose con­ The risk in subsequent time steps of the scenario is unknown and
straints on the dynamic aspects of the testing scenario. In this study, the cannot be iteratively calculated using intuitive indicators. So, a novel
parameters determined by the ODD are represented as θ. The second SM and a real-time risk evaluation model are integrated to identify the
category involves dynamic elements that cannot be universally defined potential risky interactions of vehicle and pedestrian.
by the ODD, as they vary depending on the specific scenario. These
include factors such as vehicle speed, pedestrian movement, and the • Surrogate Model
relative distance between pedestrians and vehicles. Since these elements
are scenario-specific, they require targeted research for their precise The construction of the SM is a crucial step in critical testing sce­
characterization. narios generation that could represent the function of the perception,
In the context of this paper, we focus on the dynamic elements of the decision-making and control of AV. The three-stage AEB systems have
scenario, particularly those that arise during pedestrian crossings. We been tested and demonstrated with better adaptability and driving
define the ODD for pedestrian crossings at signal-controlled comfort, while reducing false alarms and collision severity (Zhou and

6
L. Yang et al. Accident Analysis and Prevention 214 (2025) 107962

time. When the vehicle passes through the conflict area first, the TTC is
calculated by Eq. (6).
( )
dlat (i) dlon (i)
TTC(i) = max , (6)
vped (i) vveh (i)

After identifying the priority participant for each conflict scenario,

TTC is computed by frame-by-frame. The minimum TTC, denoted as
TTCmin , is extracted from each interaction set by Eq. (7), representing the
highest severity of the interaction.
TTCmin = min(TTC(i) ) (7)
When the current TTCmin is less than τFCW , the FCW is activated, and
Fig. 5. Braking strategy of three-stage AEB models.
the three-stage AEB selects different deceleration rates for collision
avoidance. The Fig. 5 is a schematic diagram of the three-stage AEB.
Wang, 2022) (Ye et al., 2024). In this paper, we use a three-stage AEB
model to be the SM. • Risk Field Model
In the three-stage AEB system, the initial action is triggered by the
Forward Collision Warning (FCW), which is activated based on the Risk assessment metrics based on the APF method have significant
stopping time varying with the AV’s speed vego . The stopping time refers advantages in the classification of risk levels with highly suitable for
to the duration from the initial application of braking abrake to a complete complex traffic scenarios. Wang et al. (Wang et al., 2015) introduced the
stop, and it is continuously compared with the current Time to Collision Driving Safety Field (DSF) model, which quantitatively describes the
(TTC). The braking process of a human-driven vehicle is generally impact of various traffic factors on driving safety. Wu et al. (Wu et al.,
divided into three phases: a uniform motion phase corresponding to 2019) further integrated a trajectory prediction module into the DSF
brake reaction time τreact , a varying deceleration phase during which model and proposed a vehicle–pedestrian risk assessment method for the
deceleration increases τbrake grow , and a uniform deceleration phase τstop , unsignalized intersection. It could reduce the braking time and increase
as shown in Eq. (3). the TTC, but it has limitations in its applicability to certain scenarios.
τFCW = τreact + τbrake + τstop (3) This study introduces a vehicle–pedestrian interaction risk field model
grow
based on the APF. It fully considers the dynamic interaction risks arising
The braking reaction time of the tested AV is minimal and can be from pedestrian crossing behaviour and accurately calculates the real-
neglected (Zhao et al., 2019). Therefore, this paper only considers the time evolution of the risk field in vehicle–pedestrian interaction
braking time in the uniform deceleration phase (Zhao et al., 2023). Eq. scenarios.
(4) represents the stopping time of AV. The DSF consists of a potential field generated by stationary objects,
vego a kinetic field formed by moving objects, and a behavioural field
τstop = (4) influenced by drivers. The safety potential SPEv,bi is defined as the scalar
abrake
potential energy possessed by the vehicle under conservative force in the
The potential conflict area between the pedestrian and the vehicle is
driving safety field, with a value of zero at an infinite distance from the
determined based on the overlap area of their trajectories, the length field source. The safety potential energy SPEv,bi of vehicle i induced by a
and width of the vehicle, as shown in Fig. 4. When the pedestrian passes
moving object b is represented as follows:
through the conflict area first, the TTC is calculated by Eq. (5).
( ) ( )1
dlat (i) + w(j) dlon (i) KRb Mb Ri Mi k3 (k3 − |vb |cosθ )1− k1 k1
TTC(i) = max , (5) SPEv,bi = • (8)
vped (i) vveh (i) (k1 − 1)|rbi | k1 − 1 k3 − |vb |

where w(j) represents the vehicle width, and i represents the current Fig. 6 illustrates the field strength Ev,bi distribution in a typical

Fig. 6. Distribution of field strength in pedestrian crossing scenario.

7
L. Yang et al. Accident Analysis and Prevention 214 (2025) 107962

Fig. 7. 4-Neighbourhood flood-fill diagram.

pedestrian crossing scenario. In a conservative vector field, the potential 5.3. Critical scenario optimization search
energy SPEv,bi is the gradient of the corresponding field strength Ev,bi ,
with the magnitude and direction of the field strength determining the 5.3.1. Auxiliary objective function
distribution of potential energy. The vehicle’s potential energy is To efficiently search for critical scenarios, this paper introduces an
influenced by the field strength generated by the crossing pedestrian, auxiliary objective function to provide a search direction. Among all the
and it varies according to parameters such as the relative distance. The frames, the risk at the most critical moment (i.e., the maximum risk
field strength expression is as follows: value) is selected as the optimization objective, as precisely defined in
Eq. (14).
KRb Mb k3 − ∇Ev,bi
Ev,bi = • ⃒⃒ ⃒ (9)
(k3 − |vb |cosθ)•|rbi |k1 Ev,bi ⃒ mRDSIv,bi (t) = maxRDSIv,bi (t) (14)
t

Where the vector vb represents the velocity of object b, while rbi To search high-frequency scenarios, the scenario with the highest
denotes the distance vector between moving object b and host vehicle i. exposure frequency is selected. The distances between this selected
The angle θb indicates the angle between rbi and vb . Constants K ( > 0) scenario and the others are then calculated to filter out the more
and k1 ( > 1) are unspecified, and k3 represents a risk coefficient. Rb and frequent scenarios. It is calculated as follows:
Ri are factors influenced by the road conditions. √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
Virtual masses Mb and Mi assess the potential driving risks associated 1 ∑dim (xi − yi )2
d(x, y) = (15)
with the object’s attributes (mass, type, speed), which mainly refers to dim i=1 U2i
the potential losses that may result from an accident. For the same type
objects, greater physical mass and speed result in a larger virtual mass, Where, x represents a search scenario, and y refers to the afore­
and thus, higher potential losses. Additionally, for equivalent physical mentioned scenario with the highest exposure frequency. dim represents
mass and speed, the potential loss from a collision depends on the type of the dimension of the decision variables that is set 3. Ui is a normalization
object involved. As an example, Mi is defined as the virtual mass: factor, which can be derived by calculating the maximum distance be­
( ) tween the scenarios in the original scenario set and scenario y. From the
Mi = mi • Ti • 1.566 × 10− 14 v6.687
i + 0.3345 (10) distribution of decision variables, U1 = 24, U2 = 29, U3 = 11 are
obtained.
The variable SPEv,bi is determined by the TPs’ position and reflects
The objective function includes both the exposure frequency distance
the distribution of risk in the space. A higher SPEv,bi value signifies
calculated from Eq. (15) and the risk proxy metric constructed by Eq.
greater risk. To account for the temporal variability, the rate of change
(14). The comprehensive objective function is derived by weighting two
of safety potential energy over time is used, which can be expressed as
metrics, as shown in Eq. (16).
follows:
( )
minJ(x) = min mRDSIv,bi (x) + ω × d(x, y) (16)
dSPEv,bi dSPEv,bi drbi x x
SPE˙ v,bi = = • (11)
dt drbi dt
5.3.2. Optimization search
A positive value of SPE˙ v,bi indicates an increase, signifying a rise in
The generation of critical scenarios is transformed into an optimi­
driving risk for vehicle i due to the moving object b. The Driving Safety
zation and search problem, aiming to find solutions that minimize the
Index (DSI) quantifies a vehicle’s driving risk by integrating its safety
objective function across various parameter combinations. After
potential energy and its rate of change over time and space, with higher
applying the SM model to the original scenario library. A scenario is
values indicating increased risk. To quantify the vehicle’s risk, the
deemed critical if its criticality surpasses a predefined threshold,
Relative Driving Safety Index (RDSI) is used in this paper.
ensuring that the critical scenario has a specified probability of occur­
DSIv,bi = αSPEv,bi + (1 − α)SPE˙ v,bi (12) rence within the scenario space. The formula for the threshold is as
follows:
DSIv,bi m
RDSIv,bi = (13) = 3.471 × 10− 4
(17)
DSIv,bi * |V(x) | >
N(X)
Where the DSIv,bi * represents the standard DSI for the vehicle i in a Where the total number of scenarios N(X) is 43,214, and m is set to
specific scenario. The dangerous scenario is classified using PET as the 15. The 4-neighbourhood flood-fill algorithm is employed to explore the
indicator, with the threshold of 3 s (Amini et al.). The final RDSI neighbourhood of critical scenarios until all adjacent scenarios are
calculated is the scenario risk. deemed non-critical, at which point the search concludes, and all critical
scenarios are identified.
As shown in Fig. 7, each local critical scenario is represented by the
central orange area. The adjacent regions—top, bottom, left, and

8
L. Yang et al. Accident Analysis and Prevention 214 (2025) 107962

Table 2
Details of the original vehicle–pedestrian interaction scenarios.
No. City Collection Number of valid Maximum vehicle speed Maximum pedestrian speed Number of Number of
Duration scenarios (m/s) (m/s) vehicles pedestrians

1 Xi’an 615 min 3679 15.7 4.0 1736 1201

2 Changchun 163 min 555 15.1 3.7 453 214
3 Chongqing 227 min 805 16.8 3.2 297 164
4 Tianjin 420 min 846 15.7 2.9 399 103

Fig. 8. Distribution of pedestrian trajectories density.

right—are evaluated individually to determine if the conditions outlined fewest valid vehicle–pedestrian interaction scenarios, indicating a lower
in Eq. (17) are met. For instance, after a single evaluation, if the top, potential for conflicts. In contrast, the Xi’an intersection has more
bottom, and right neighbouring regions satisfy the conditions, they are dispersed pedestrian trajectories, suggesting a higher potential for
marked and filled with a new colour to denote critical scenarios. conflicts, as seen by the largest number of valid interaction scenarios.
Through this recursive process, all critical scenarios within the scene The trajectory distributions at Chongqing and Tianjin are similar, with
space are identified and appropriately marked. most pedestrian paths within the crosswalk and a few outside, resulting
in similar counts of extracted scenarios, placing them in the middle
6. Experiment and result analysis range.

6.1. Original scenarios library for vehicle–pedestrian interaction 6.2. Distribution of decision variables

As shown in Table 2, a total of 5,885 original vehicle–pedestrian The distribution of the vehicle’s longitudinal speed at the moment of
interaction scenarios were extracted from the SIND dataset that initial encounter is shown in Fig. 9. In urban traffic environments,
contribute to the original vehicle–pedestrian interaction scenario li­ typical low-speed ranges of vehicles are between 10–30 km/h, while the
brary. Fig. 8 illustrates the density distribution of complete pedestrian normal driving range spans 30–60 km/h. Fig. 9(b) shows the distribu­
trajectories within the valid vehicle–pedestrian interaction scenarios tion of pedestrian speeds at the moment of initial encounter. A few of
extracted from each intersection. It is important to note that a complete speed values exceeds 3 m/s, which is at the upper end of pedestrian
pedestrian trajectory refers to the entire set of data for a pedestrian speeds, likely due to pedestrians accelerating to cross before the light
within the dataset, rather than partial trajectory data from the interac­ changes or engaging in jaywalking. Results from Fig. 9(c) and 9(d)
tion process. Table 2 shows that the Changchun intersection has the reveal that the longitudinal distance and lateral relative distance

9
L. Yang et al. Accident Analysis and Prevention 214 (2025) 107962

Fig. 9. Statistical distribution of decision variables in original scenarios.

between the vehicle and pedestrian at the moment of initial encounter 1.35m/s2 ), and the third full braking stage d3three is at the 95th percentile
do not follow a normal distribution. It means that this is the high (d3three = 2m/s2 ). The stopping time for each braking stage is computed
randomness and uncertainty in the timing of pedestrians arriving at the using Eq. (4), and braking pressure is applied if the TTC falls below this
crosswalk and choosing their crossing moment. It is influenced by in­ stopping time.
dividual pedestrian behaviour and real-time traffic conditions, such The parameters setting of the risk model are shown in Table 3. The
factors exhibit limited regularity, significantly impacting the distribu­ selection of each parameter is guided by the approach of Chen, Wang,
tion of vehicle–pedestrian relative distances. et al. (Chen et al., 2022), who calibrated parameters such as virtual mass
and road condition influence factors using the same model employed in
6.3. Scenario exposure frequency of vehicle–pedestrian interaction this study, drawing on the Chinese Road Traffic Safety dataset. The
validity of these calibrated parameters was validated through vehicle
As shown in Fig. 10, coordinate axes are established based on relative experiments.
lateral distance dlat , relative longitudinal distance dlon , and pedestrian- The 90th percentile value is selected as the risk threshold for vehi­
vehicle speed difference diff v− p . The legend represents the exposure cle–pedestrian interaction. When the risk value of any frame during the
frequency of corresponding scenarios. The colour closer to red indicates interaction process in a given scenario exceeds this threshold, the sce­
the higher exposure frequency which means a higher likelihood of nario is deemed to present significant risk which is classified as a
occurrence in real-world traffic environments. As shown in Fig. 10(a), it dangerous scenario. Ultimately, 173 dangerous scenarios were extracted
can be seen that in the original scenarios, the dlat is mainly concentrated from the original scenario library. The specific distribution of dangerous
in 10 m, while the dlon is more dispersed. This may be because the dlat can scenarios is shown in Fig. 12 with yellow areas.
be approximated as the distance from the edge of a crosswalk to the To validate the effectiveness of the risk model proposed in this study,
middle of a lane, which is relatively fixed. The pedestrian arrival at the the inverse of TTC, denoted as TTCi , is analysed as a reference value for
crosswalk edge during vehicle travel is random will leads to the RDSI, suggesting that TTCi is directly proportional to the level of risk.
dispersed distribution of dlon . The three-dimensional exposure frequency Fig. 13 and Fig. 14(a) present a comparison between the proposed risk
is shown in Fig. 10(c). Fig. 10(d) shows that the diff v− p is mainly within model and TTCi , and Fig.14(b) shows the TTC of pedestrian and vehicle
6 m/s. At signalized intersections, factors such as congestion and in each interaction scenario. In each frame, the reciprocal of the larger
yielding will cause vehicles to idle or move slowly, resulting in a rela­ TTC value between them is taken as the TTCi value. As shown in Fig. 13,
tively lower speed difference with pedestrians. the risk model curve is highly consistent with the TTCi curve. It means
that the proposed risk model could effectively reflect the potential
6.4. Risk analysis of vehicle–pedestrian interaction scenarios interaction risks. As demonstrated in Fig. 14, the TTC metric is insen­
sitive to risks posed by laterally moving objects, leading to a delayed
Fig. 11(a) shows the activation of the first braking stage when TTC response in pedestrian crossing scenarios compared to the proposed risk
drops below FCW, where the vehicle applies braking force d1three . In this model. It indicates that the proposed risk model has superior capability
study, the three stages correspond to three levels of braking pressure, in assessing lateral risks compared to TTC.
representing percentiles of the maximum deceleration observed in the
original scenarios. The deceleration distribution is shown in Fig. 11(b). 6.5. Pedestrian critical testing scenarios
The first-stage braking pressure d1three is at the 50th percentile (d1three =
0.9m/s2 ), the second-stage d2three is at the 75th percentile (d2three = After optimizing search in the original scenario library, a total of 50

10
L. Yang et al. Accident Analysis and Prevention 214 (2025) 107962

Fig. 10. Distribution of exposure frequency in original scenarios.

Fig. 11. Activation and braking parameters of the three-stage AEB model.

Table 3
Parameters setting of the vehicle–pedestrian interaction risk model.
Parameter Value Parameter Value Parameter Value Parameter Value

k 0.1 k1 1.5 Rb 1 Ri 1
k3 160 α 0.9 Tb 1.136 Ti 1

11
L. Yang et al. Accident Analysis and Prevention 214 (2025) 107962

Fig. 12. Distribution of exposure frequency in hazardous scenarios.

Fig. 13. Comparative analysis of vehicle–pedestrian interaction risk and TTCi trends (Scenario with Minimum TTC = 1.57 s).

Fig. 14. Comparative analysis of vehicle–pedestrian interaction risk and TTCi trends (Scenario with Minimum TTC = 1.19 s).

12
L. Yang et al. Accident Analysis and Prevention 214 (2025) 107962

Fig. 15. Distribution of exposure frequency in critical scenarios.

Fig. 16. Partial pedestrian behaviours in critical scenarios.

critical scenarios were generated, and their distribution presented is Behaviours such as high-speed zigzagging, jaywalking, and interactions
shown in Fig. 15. Fig. 16 illustrates a selection of distinct pedestrian with turning vehicles could test the AV’s ability to anticipate unpre­
movement characteristics extracted from the ultimately identified crit­ dictable movements. High-speed crossings with evasive actions or sud­
ical scenarios. Each identified critical scenario is analysed individually den stops challenge the AV’s real-time responsiveness and adaptive
in Fig. 17, and detail classification of pedestrian crossing characteristics decision-making. Non-standard crossings, such as diagonal movements
is shown in Table 4. or rapid direction changes, would affect the AV’s environmental
As shown in Table 4, the experiment categorizes pedestrian behav­ perception and its capacity to adapt to dynamic conditions.
iours into 7 types based on the 50 critical test scenarios generated by the Fig. 17 illustrates the variation in risk quantification values for
proposed method which is crucial for AV pedestrian testing. These sce­ critical scenarios with respect to pedestrian speed, vehicle speed, and
narios are typical and high-risk at signalized intersections that are their relative distance. As shown Fig. 17(a), the variation in risk for
challenges for the perception, prediction, and decision-making of AV. critical scenarios is analysed with respect to the pedestrian speed,

13
L. Yang et al. Accident Analysis and Prevention 214 (2025) 107962

Fig. 17. Risk associated with different speeds and distances in critical scenarios.

14
L. Yang et al. Accident Analysis and Prevention 214 (2025) 107962

Fig. 17. (continued).

To verify the effectiveness of the proposed method, the minimum

Table 4 TTC and PET are selected as baselines for comparison. The minimum
Description of pedestrian behaviour in critical scenarios. TTC refers to the lowest TTC value throughout the interaction process,
No. Description representing the most hazardous moment of the entire interaction. The
1 Pedestrian crosses the road at a high speed with frequent changes in path final comparative analysis results are shown in Fig. 18. In this study,
direction, repeatedly zigzagging across the street. scenarios with a PET greater than 3 s and a TTC greater than 2.72 s are
2 Pedestrian crosses the road at a high speed while jaywalking or during the end classified as low-risk scenarios (Sheykhfard et al., 2023). Among the
of a green light.
critical scenarios generated by the proposed framework, 30 scenarios
3 Pedestrian crosses at a high speed while interacting with turning vehicles.
4 Pedestrian crosses at a high speed, then stops or reduces speed to yield to have a PET greater than 3 s, and 22 scenarios have a minimum TTC
passing vehicles before resuming. greater than 2.72 s. The distribution of critical scenarios is primarily
5 Pedestrian crosses at a high speed, taking evasive action around vehicles in the concentrated within a vehicle–pedestrian distance of 15 m and a speed
conflict area. difference of 4 m/s, with a small number of scenarios exhibiting a
6 Pedestrian crosses outside of crosswalks or diagonally at intersections.
7 Pedestrian crosses and then rapidly turns back with high speed.
scattered distribution, where both the vehicle–pedestrian distance and
speed difference gradually increase.
This paper further analysis and validation of the aforementioned
vehicle speed, and their relative distance. It illustrates diverse patterns conclusions by three critical scenarios. The PET and Minimum TTC
in pedestrian and vehicle interactions, highlighting instances where values for the three scenarios are as follows: for the first scenario, PET =
high pedestrian speeds and shorter pedestrian-vehicle distances corre­ 28.9 s and Minimum TTC = 3.9 s; for the second scenario, PET = 3.4 s
spond to elevated risk levels. The trajectory trends reveal significant and Minimum TTC = 10.9 s; and for the third scenario, PET = 2.4 s and
dependencies between these variables, with the colour gradient further Minimum TTC = 1.7 s. The corresponding figures for these scenarios are
emphasizing the intensity of the risks under different conditions. It un­ shown in Fig. 19, Fig. 20, and Fig. 21, respectively.
derscores the dynamic nature of critical scenarios and their contribution In the analysis of scenarios with large PET and small minimum TTC,
to the understanding of pedestrian-vehicle interactions in autonomous it is evident that the larger PET value arises from the pedestrian waiting
driving testing. Fig. 17(b) shows the three-dimensional aggregated dis­ for an extended period due to the red light. Simultaneously, the vehicle
tribution. It can be observed that higher risk is generally associated with passes through the conflict area shortly after the initial encounter with
faster pedestrian speeds and shorter distances between pedestrians and the pedestrian, contributing to a larger PET. However, at the moment of
vehicles (Lanzaro et al., 2023). initial interaction, the pedestrian and vehicle are in close proximity,

Fig. 18. Comparison of PET and TTC in critical scenarios.

15
L. Yang et al. Accident Analysis and Prevention 214 (2025) 107962

Fig. 19. A critical scenario classified as safe by PET but hazardous by Minimum TTC.

Fig. 20. A critical scenario classified as hazardous by PET but safe by Minimum TTC.

Fig. 21. A critical scenario classified as hazardous by both PET and Minimum TTC.

resulting in a smaller minimum TTC value. The vehicle decelerates upon vital for assessing the vehicle’s ability to evaluate and react to dynamic,
detecting the pedestrian’s intent to cross, resuming acceleration only real-world risks.
after confirming the pedestrian has stopped waiting. This indicates that In the analysis of the scenario shown in Fig. 20, where PET is small
the vehicle recognized a certain level of risk in the interaction. Despite and the minimum TTC is large, it is observed that the small PET results
the larger PET and smaller TTC of this scenario, it is still highly relevant from the vehicle accelerating and quickly leaving the conflict area after
for testing, as they demonstrate situations where the vehicle must the pedestrian crossed, while the initial distance between the pedestrian
respond to potential interaction risks, even when the pedestrian’s ac­ and the vehicle was relatively short. The large minimum TTC, however,
tions do not immediately indicate imminent danger. Such scenarios are is due to the vehicle decelerating at the moment of initial encounter and

16
L. Yang et al. Accident Analysis and Prevention 214 (2025) 107962

waiting at a red light while the pedestrian crossed. The vehicle only significant uncertainty and randomness. The critical testing scenarios
began to accelerate and leave the conflict area shortly after the pedes­ encompass a broader range of pedestrian characteristics, such as
trian had cleared it. Despite the vehicle waiting for the pedestrian to detouring, decelerating to avoid vehicles, and jaywalking. Additionally,
cross due to the red light, it started accelerating before the pedestrian 30 of the critical scenarios feature pedestrians with maximum speeds
had fully cleared the conflict area. This suggests that, even though the exceeding 1 m/s, maintaining a high crossing speed. This provides
vehicle was temporarily halted, the scenario still carries a certain level robust support for testing AV in diverse and dynamic pedestrian
of risk, as the vehicle’s premature acceleration introduces potential interactions.
conflict if the pedestrian had not completely cleared the area. (3) Among the 50 generated high-value testing scenarios, a key
In the critical scenario where both PET and minimum TTC are small, finding is that higher pedestrian speeds and shorter relative distances
as shown in Fig. 21, it can be observed that the pedestrian maintained a between pedestrians and vehicles are strongly correlated with elevated
high speed throughout the crossing, and the vehicle also approached risk levels. This insight provides a valuable foundation for the iterative
with a high speed at the initial encounter. According to the SM calcu­ optimization of decision-making and control algorithms in AV.
lation, the three-stage AEB was activated during this process. The The limitations of this study arise from the fact that the descriptive
vehicle detected the pedestrian and began braking when the current TTC features of vehicle–pedestrian interaction scenarios are currently
became shorter than the required stopping time, eventually decelerating limited to relative distance and relative speed, without considering more
and leaving the conflict area, thus marking the end of the interaction. In detailed real-world characteristics such as pedestrian gender, age, and
this scenario, both the pedestrian and the vehicle maintained high distracted behaviours. At present, the SIND dataset does not include
speeds, and the distance between them remained relatively short as the information such as the gender and age of pedestrians, but future
interaction progressed. As a result, PET, TTC, and the risk model pro­ research will focus on refining the dataset to incorporate these addi­
posed in this paper all identified this scenario as a high-risk interaction, tional factors. This will help create a more comprehensive library of
underlining the potential danger due to the short distance and high pedestrian testing scenarios for AV.
speeds involved.
By the analysis of the critical scenarios above, it is clear that the risk CRediT authorship contribution statement
model proposed in this paper effectively addresses the limitations of
relying on a single SSM. By incorporating multiple factors, the model Lan Yang: Writing – original draft, Validation, Project administra­
successfully identifies interaction scenarios with significant conflicts, tion. Songyan Liu: Writing – original draft, Software, Data curation.
offering a more accurate and comprehensive assessment of vehi­ Shuo Feng: Methodology, Conceptualization. Hong Wang: Supervi­
cle–pedestrian interaction risks. sion, Data curation. Xiangmo Zhao: Writing – review & editing,
Conceptualization. Guangyue Qu: Visualization, Conceptualization.
7. Conclusion Shan Fang: Investigation, Formal analysis.

To further assess the safety and adaptability of AV in real-world Declaration of competing interest
traffic environments, this study builds upon the Chinese traffic micro-
dataset SIND and proposes a method for generating critical pedestrian The authors declare that they have no known competing financial
testing scenarios. The objective is to provide typical, high-value vehi­ interests or personal relationships that could have appeared to influence
cle–pedestrian interaction scenarios for AV testing. The key findings of the work reported in this paper.
this paper are as follows:
(1) A risk level evaluation method for vehicle–pedestrian in­ Acknowledgments
teractions based on the APF approach is proposed. Experiments reveal
that in the 50 critical scenarios generated by this method, the PET metric This study was jointly supported by the National Natural Science
identifies 20 risk scenarios (40 % of the total), while the TTC metric Foundation of China (No. 52472446 and No. 62473224), the National
identifies 28 (56 % of the total). In contrast, the proposed method suc­ Key Research and Development Program of China (No.
cessfully captures all critical scenarios, showcasing superior recognition 2021YFB2501205), and the Shaanxi Province Merit-based Funding
capability. The findings indicate that the proposed method is more project for Science and Technology Activities of Overseas Educated
comprehensive and accurate in identifying risks within complex vehi­ Personnel (No. 2023001). The authors appreciate the support of data
cle–pedestrian interaction scenarios. from the SOTIF Workgroups of the China Alliance for Intelligent Internet
(2) Pedestrian behaviour at signalized scenarios often involves Connected Vehicle Industry Innovation (CAICV).

Appendix A

The process of generating critical scenarios is presented in Table 5. This algorithm generates critical scenarios within a scenario space using
optimization and search methods. First, the set of critical scenarios, C, is initialized as an empty set. Then, an objective function J(x) is designed to
assess the criticality of scenarios and serves as the basis for optimization in the subsequent steps. Next, an appropriate optimization algorithm is
selected to minimize J(x), thereby obtaining a candidate scenario xʹ. Once the candidate scenarios are obtained, the algorithm filters those that satisfy
the conditions specified in Eq. (17), marking them as local critical scenarios x* . These local critical scenarios meet a certain criticality threshold,
representing scenarios with higher risks or special importance. The algorithm then explores and fills the surrounding areas of these local critical
scenarios x* , recursively generating other critical scenarios connected to them. In the end, all scenarios that meet the criticality threshold are collected
in the set C .
The key idea of this process is to use an optimization algorithm combined with the criticality function to progressively generate and identify critical
scenarios in the scenario space. Through the recursive filling process, it ensures that all critical scenarios are accurately generated and marked.

17
L. Yang et al. Accident Analysis and Prevention 214 (2025) 107962

Table 5
The process of critical scenario generation.
Critical Scenario Generation

Input: Scenario parameter space X, criticality function V(x)

Output: Set of critical scenarios C
1. Initialize C as empty
2. Design the objective function J(x)
3. Select an optimization algorithm to minimize J(x), and obtain the candidate scenario xʹ
4. Filter candidate scenarios that satisfy Eq. (17) as the identified local critical scenarios x*
m
5. Explore and fill the surrounding area of x* , obtaining all x* , C = {x ∈ X : V(x) > }
N(X)
return C

Data availability Evaluation Framework (ISO Standard No. 34502), 2022. https://www.iso.org/
standard/78951.html.
International Organization for Standardization (ISO), Road Vehicles—Safety Evaluation
The data that has been used is confidential. Framework for Automated Driving Systems—Part 1: Evaluation Framework (ISO
Standard No. 34501), 2022. https://www.iso.org/standard/78950.html.
International Organization for Standardization (ISO), Road Vehicles—Safety of the
References Intended Functionality (ISO Standard No. 21448), 2022. https://www.iso.org/
standard/77490.html.
Abeysirigoonawardena, Y., Shkurti, F., Dudek, G., 2021. Generating adversarial driving Klück, F., Zimmermann, M., Wotawa, F., Nica, M., 2019. Genetic algorithm-based test
scenarios in high-fidelity simulators. In: 2019 International Conference on Robotics parameter optimization for ADAS system testing. In: 2019 IEEE 19th International
and Automation. ICRA, pp. 8271–8277. https://doi.org/10.1109/ Conference on Software Quality, Reliability and Security. QRS, pp. 418–425.
ICRA.2019.8793740. https://doi.org/10.1109/QRS.2019.00058.
Amini, R.E., Yang, K., Antoniou, C., 2022. Development of a conflict risk evaluation Krajewski, R., Moers, T., Nerger, D., Eckstein, L., 2018. Data-driven maneuver modeling
model to assess pedestrian safety in interaction with vehicles. Accid. Anal. Prev. 175, using generative adversarial networks and variational autoencoders for safety
306–324. https://doi.org/10.1016/j.aap.2022.106773. validation of highly automated vehicles. In: 2018 21st International Conference on
Angulo, A.V., Robartes, E., Guo, X., Chen, T.D., Heydarian, A., Smith, B.L., 2023. Intelligent Transportation Systems. ITSC, pp. 2383–2390. https://doi.org/10.1109/
Demonstration of virtual reality simulation as a tool for understanding and ITSC.2018.8569971.
evaluating pedestrian safety and perception at midblock crossings. Transp. Res. Kuang, S., Liu, Y., Wang, X., Wu, X., Wei, Y., 2024. Harnessing multimodal large
Interdiscip. Perspect. 20, 100844. https://doi.org/10.1016/j.trip.2023.100844. language models for traffic knowledge graph generation and decision-making.
Baby, T., Şahin Ippoliti, H., Wintersberger, P., Zhang, Y., Yoon, S.H., Lee, J., Lee, S.C., Commun. Transp. Res. 4, 100146. https://doi.org/10.1016/j.commtr.2024.100146.
2024. Development and classification of autonomous vehicle’s ambiguous driving Kwon, J.-H., Kim, J., Kim, S., Cho, G.-H., 2022. Pedestrians’ safety perception and
scenario. Accid. Anal. Prev. 200, 107501. https://doi.org/10.1016/j. crossing behaviors in narrow urban streets: an experimental study using immersive
aap.2024.107501. virtual reality technology. Accid. Anal. Prev. 174, 106757. https://doi.org/10.1016/
Bagschik, G., Menzel, T., Maurer, M. 2018. Ontology based scene creation for the j.aap.2022.106757.
development of automated vehicles. 2018 IEEE Intelligent Vehicles Symposium (IV). Lanzaro, G., Sayed, T., Fu, C., 2023. A comparison of pedestrian behavior in interactions
IEEE, 2018: 1813-1820. 10.1109/IVS.2018.8500632. with autonomous and human-driven vehicles: an extreme value theory approach.
Bruggner, D., Hegde, A., Acerbo, F.S., Gulati, D., Son, T.D., 2021. Model in the loop Transp. Res. Part f: Traffic Psychol. Behav. 99, 1–18. https://doi.org/10.1016/j.
testing and validation of embedded autonomous driving algorithms. 2021 IEEE trf.2023.10.006.
Intelligent Vehicles Symposium. IV, pp. 136-141. 10.1109/IV48863.2021.9575530. Lenard, J., Welsh, R., Danton, R., 2018. Time-to-collision analysis of pedestrian and
Cai, J., Deng, W., Guang, H., Wang, Y., Li, J., Ding, J., 2022. A survey on data-driven pedal-cycle accidents for the development of autonomous emergency braking
scenario generation for automated vehicle testing. Machines 10 (11), 1101. https:// systems. Accid. Anal. Prev. 115, 128–136. https://doi.org/10.1016/j.
doi.org/10.3390/machines10111101. aap.2018.02.028.
Cao, Y., Zhou, H., Xiao, L., Wang, Y., Li, P., Tan, Z., Qu, X., Luo, J., 2020. Analysis of Li, Y., Yang, Z., Jin, J., Wu, D., 2024. Exploring master scenarios for autonomous driving
pedestrian-vehicle collision accident characteristics based on video data from the tests from police-reported historical crash data using an adaptive search sampling
NAIS database. J. Automob. Saf. Energy 11 (01), 44–52. https://doi.org/10.3969/j. framework. Accid. Anal. Prev. 205, 107688. https://doi.org/10.1016/j.
issn.1674-8484.2020.01.004. aap.2024.107688.
Chen, H., Wang, H., Li, W., Bai, X., Chen, J., Li, C., Shi, Q., Sun, J., 2022. Risk assessment Li, Y., Bai, F., Lyu, C., Qu, X., Liu, Y., 2025. A systematic review of generative adversarial
of expected functional safety scenarios based on the driving safety field theory. networks for traffic state prediction: overview, taxonomy, and future prospects. Inf.
Automot. Eng. 44 (11), 1636–1646. https://doi.org/10.19562/j.chinasae. Fusion., 102915 https://doi.org/10.1016/j.inffus.2024.102915.
qcgc.2022.11.002. Lin, H., Liu, Y., Wang, L., Qu, X., 2025. Big data-driven advancements and future
China New Car Assessment Programme (C-NCAP), (2024). https://www.c-ncap.org.cn directions in vehicle perception technologies: from autonomous driving to modular
/article-detail/1747900203303780353?type=2. buses. IEEE Trans. Big Data. https://doi.org/10.1109/TBDATA.2025.3527208.
European New Car Assessment Programme (Euro NCAP), (2023). Test Protocol—AEB Liu, Y., He, J., Liu, W., Zhu, X., Li, L., Jiang, L., 2012. Research on pedestrian test
Vulnerable Road User (VRU) Systems. https://cdn.euroncap.com/media/75436/e scenarios for automatic emergency braking system. Automob. Technol. 3, 35–39.
uro-ncap-aeb-lss-vru-test-protocol-v43.pdf. https://doi.org/10.3969/j.issn.1000-3703.2014.03.009.
Fei, Y., Shi, P., Liu, Y., Wang, L., 2024. Critical roles of control engineering in the Ma, J., Tian, Y., Zhao, X., Sun, J., Wang, H., 2021. Two-stage test scenarios for
development of intelligent and connected vehicles. J. Intell. Connect. Vehicles 7 (2), interaction of vehicles and pedestrians based on naturalistic driving data mining.
79–85. https://doi.org/10.26599/JICV.2023.9210040. China. J. Highw. Transp. 35 (3), 139–152. https://doi.org/10.19562/j.chinasae.
Feng, S., Feng, Y., Yu, C., Zhang, Y., Liu, H.X., 2021. Testing scenario library generation qcgc.2021.04.019.
for connected and automated vehicles, part I: methodology. IEEE Trans. Intell. Mathieu, E., Rainforth, T., Siddharth, N., Teh, Y.W., 2019. Disentangling
Transp. Syst. 22 (3), 1573–1582. https://doi.org/10.1109/TITS.2020.2972211. disentanglement in variational autoencoders. Proceedings of the 36th International
Feng, S., Feng, Y., Sun, H., Bao, S., Zhang, Y., Liu, H.X., 2021. Testing scenario library Conference on Machine Learning, 97. 10.48550/arXiv.1812.02833.
generation for connected and automated vehicles, part II: case studies. IEEE Trans. Muktadir, G.M., Whitehead, J., 2024. Adaptive pedestrian agent modeling for scenario-
Intell. Transp. Syst. 22 (9), 5635–5647. https://doi.org/10.1109/ based testing of autonomous vehicles through behavior retargeting. In: 2024 IEEE
TITS.2020.2988309. International Conference on Robotics and Automation. ICRA, pp. 16595–16602.
Standardization Administration of China. Intelligent and connected vehicles — Road test https://doi.org/10.1109/ICRA57147.2024.10610078.
methods and requirements for autonomous driving functions (GB/T 44719). http:// Qu, X., Lin, H., Liu, Y., 2023. Envisioning the future of transportation: Inspiration of
www.catarc.org.cn/upload/202012/29/202012291014043001.pdf. ChatGPT and large models. Commun. Transp. Res. 3, 100103. https://doi.org/
Global status report on road safety 2023, World Health Organization (WHO). https 10.1016/j.commtr.2023.100103.
://www.who.int/teams/social-determinants-of-health/safety-and-mobility/globa Rasouli, A., Tsotsos, J.K., 2020. Autonomous vehicles that interact with pedestrians: a
l-status-report-on-road-safety-2023. survey of theory and practice. IEEE Trans. Intell. Transp. Syst. 21 (3), 900–918.
Hu, L., Yi, P., Huang, J., Zhang, X., Lei, Z., 2018. Research on two-wheeler test scenarios https://doi.org/10.1109/TITS.2019.2901817.
for automatic emergency braking system based on real accident cases. Automob. Rezwana, S., Lownes, N., 2024. Interactions and behaviors of pedestrians with
Eng. 40 (12), 1435–1446. https://doi.org/10.19562/j.chinasae.qcgc.2018.012.009. autonomous vehicles: a synthesis. Futur. Transp. 4 (3), 722–745. https://doi.org/
International Organization for Standardization (ISO), Road Vehicles—Functional Safety 10.3390/futuretransp4030034.
(ISO Standard No. 26262), 2018. https://www.iso.org/standard/68383.html. Sauerbier, J., Bock, J., Weber, H., Eckstein, L., 2019. Definition of scenarios for safety
International Organization for Standardization (ISO), Road Vehicles—Safety Evaluation validation of automated driving functions. ATZ Worldwide 121 (1), 42–45. https://
Framework for Automated Driving Systems—Part 2: Scenario-Based Safety doi.org/10.1007/s38311-018-0197-2.

18
L. Yang et al. Accident Analysis and Prevention 214 (2025) 107962

Sengupta, A., Guler, S.I., Gayah, V.V., Warchol, S., 2024. Evaluating the reliability of Xu, Y., Li, K., Wu, X., Li, X., Li, J., Liu, H.X., Guo, J., 2022. SIND: A drone dataset at
automatically generated pedestrian and bicycle crash surrogates. Accid. Anal. Prev. signalized intersections in China, in Proc. 2022 IEEE 25th Int. Conf. Intell. Transp.
203, 107614. https://doi.org/10.1016/j.aap.2024.107614. Syst. ITSC, pp. 2471-2478. 10.1109/ITSC55140.2022.9921959.
Shaon, M.R.R., Schneider, R.J., Qin, X., He, Z., Sanatizadeh, A., Flanagan, M.D. 2018. Xu, X., Hu, W., Dong, H., Wang, Y., Xiao, L., Li, P., 2021. A review of key technologies for
Exploration of pedestrian assertiveness and its association with driver yielding constructing autonomous vehicle test scenarios. Automob. Eng. 43 (4), 610–619.
behavior at uncontrolled crosswalks. 2672. 10.1177/0361198118790645. https://doi.org/10.19562/j.chinasae.qcgc.2021.04.019.
Sheykhfard, A., Haghighi, F., Kavianpour, S., Das, S., Farahani, P.S., Fountas, G., 2023. Yang, Z., Gong, Z., Zhang, Q., Wang, J., 2022. Analysis of pedestrian-related crossing
Risk assessment of pedestrian red-light violation behavior using surrogate. IATSS behavior at intersections: a Latent Dirichlet Allocation approach. Int. J. Transp. Sci.
Res. 47 (4), 514–525. https://doi.org/10.1016/j.iatssr.2023.11.003. Technol. 12 (4), 1052–1063. https://doi.org/10.1016/j.ijtst.2022.12.003.
Sohrabi, S., Khodadadi, A., Mousavi, S.M., Dadashova, B., Lord, D., 2021. Quantifying Yang, L., Yuan, J., Zhao, X., Fang, S., He, Z., Zhan, J., 2023. SceGAN: a method for
the automated vehicle safety performance: a scoping review of the literature, generating autonomous vehicle cut-in scenarios on highways based on deep
evaluation of methods, and directions for future research. Accid. Anal. Prev. 152, learning. J. Intell. Connect. Vehicles 6 (4), 264–274. https://doi.org/10.26599/
106003. https://doi.org/10.1016/j.aap.2021.106003. JICV.2023.9210023.
Spooner, J., Palade, V., Cheah, M., Balasubramanian, V., 2021. Generation of pedestrian Ye, C., Wang, X., Morris, A., Ying, Z., 2024. Pedestrian crash causation analysis and
crossing scenarios using ped-cross generative adversarial network. Appl. Sci. 11 (2), active safety system calibration. Accid. Anal. Prev. 195, 107404. https://doi.org/
471. https://doi.org/10.3390/app11020471. 10.1016/j.aap.2024.107404.
Su, J., Chen, J., Wang, H., Chen, W., Wang, K., 2017. Pedestrian traffic conflict typical Yu, T., Long, Z., Cui, Z., Xu, L., Li, L., Zhang, C., 2023. Pedestrian simulation by learning
scene extraction and analysis based on dangerous conditions in China. Traffic from online human demonstration with VR headset. In: 2023 IEEE 26th International
Transp. Acad. Ed. 1, 209–214. https://doi.org/10.3969/j.issn.1671-3400.2017. Conference on Intelligent Transportation Systems. ITSC, pp. 3680–3685. https://doi.
z1.049. org/10.1109/ITSC57777.2023.10422058.
Sui B., Zhou S., Zhao x., Lubbe N., 2017. An Overview of Car-To-Two-Wheeler Accidents Zhan, W., Sun, L., Wang, D., Shi, H., Clauss, A., Naumann, M., Tomizuka, M., 2019.
in China: Guidance for AEB Assessment. The 25th International Technical INTERACTION Dataset: An INTERnational, Adversarial and Cooperative moTION
Conference on the Enhanced Safety of Vehicles. https://www.researchgate.net/pu Dataset in Interactive Driving Scenarios with Semantic Maps. arXiv preprint arXiv:
blication/317549331. 1910.03088, 2019. https://arxiv.org/abs/1910.03088.
Sun, L., Huang, S., Zheng, C., Bai, T., Hu, Z., 2023. Test case generation for autonomous Zhang, Y., Wang, S., Chen, B., Cao, J., Huang, Z., 2021. TrafficGAN: Network-scale deep
driving based on improved genetic algorithm. In: 2023 IEEE 23rd International traffic prediction with generative adversarial nets. IEEE Trans. Intell. Transp. Syst.
Conference on Software Quality, Reliability, and Security. QRS, pp. 272–278. 22 (1), 219–230. https://doi.org/10.1109/TITS.2019.2955794.
https://doi.org/10.1109/QRS60937.2023.00035. Zhang, P., Zhu, B., Zhao, J., Fan, T., Sun, Y., 2023. Safety evaluation method in multi-
Tian, K., Tzigkarias, A., Wei, C., Lee, Y.M., Holmes, C., Leonetti, M., Merat, N., logical scenarios for automated vehicles based on naturalistic driving trajectory.
Romano, R., Markkula, G., 2023. Deceleration parameters as implicit Accid. Anal. Prev. 180, 106926. https://doi.org/10.1016/j.aap.2022.106926.
communication signals for pedestrians’ crossing decisions and estimations of Zhao, D., Huang, X., Peng, H., Lam, H., LeBlanc, D.J., 2018. Accelerated evaluation of
automated vehicle behaviour. Accid. Anal. Prev. 190, 107173. https://doi.org/ automated vehicles in car-following maneuvers. IEEE Trans. Intell. Transp. Syst. 19
10.1016/j.aap.2023.107173. (3), 733–744. https://doi.org/10.1109/TITS.2017.2701846.
Tran, T.T.M., Parker, C., 2024. Advancing VR simulators for autonomous Zhao, X., Jing, S., Hu, F., Liu, R., Khattak, A.J., 2019. DSRC-based rear-end collision
vehicle–pedestrian interactions: a focus on multi-entity scenarios. Transp. Res. Part F warning system—An error-component safety distance model and field test. Transp.
Traffic Psychol. Behav. 107, 775–790. https://doi.org/10.1016/j.trf.2024.10.006. Res. Part C: Emerg. Technol. 107, 92–104. https://doi.org/10.1016/j.
Ulbrich S, Menzel T, Reschka A, Schuldt F, Maurer M., 2015. Defining and substantiating trc.2019.08.002.
the terms scene, situation, and scenario for automated driving. IEEE 18th Zhao, X., Zhao, J., Zhao, Y., Sun, J., 2023. Hazardous lane-change scenario
international conference on intelligent transportation systems. IEEE, 2015: 982-988. generalization generation method and system for autonomous driving testing. Acta
10.1109/ITSC.2015.164. Autom. Sin. 49 (10), 2211–2223. https://doi.org/10.16383/j.aas.c220772.
United Nations Economic Commission for Europe (UNECE). https://www.unece.org. Zhou, R., Huang, H., Lee, J., Huang, X., Chen, J., Zhou, H., 2023. Identifying typical pre-
Wang, J.Q., Wu, J., Li, Y., 2015. The driving safety field based on driver-vehicle-road crash scenarios based on in-depth crash data with deep embedded clustering for
interactions. IEEE Trans. Intell. Transp. Syst. 16 (4), 2203–2214. https://doi.org/ autonomous vehicle safety testing. Accid. Anal. Prev. 191, 107218. https://doi.org/
10.1109/TITS.2015.2401837. 10.1016/j.aap.2023.107218.
Wang, J.Q., Wu, J., Zheng, X., Li, Y., Liang, C., 2016. Driving safety field theory modeling Zhou, W., Wang, X., 2022. Calibrating and comparing autonomous braking systems in
and its application in pre-collision warning system. Transp. Res. Part C: Emerg. motorized-to-non-motorized-vehicle conflict scenarios. IEEE Trans. Intell. Transp.
Technol. 72, 2016. https://doi.org/10.1016/j.trc.2016.10.003. Syst. 23 (11), 20636–20651. https://doi.org/10.1109/TITS.2022.3202510.
Wang, X., Ye, C., Quddus, M., Morris, A., 2023. Pedestrian safety in an automated driving Zhou, J., Wang, L., Wang, X., 2024. Online adaptive generation of critical boundary
environment: calibrating and evaluating the responsibility-sensitive safety model. scenarios for evaluation of autonomous vehicles. IEEE Trans. Intell. Transp. Syst. 24,
Accid. Anal. Prev. 192, 107265. https://doi.org/10.1016/j.aap.2023.107265. 1–17. https://doi.org/10.1109/TITS.2023.3248121.
Wei, Z., Zhou, H., Zhou, R., 2024. Risk and complexity assessment of autonomous vehicle Zhu, B., Huang, Y., Zhao, J., Zhao, W., Zhang, P., Song, D., 2024. Camera-in-the-loop
testing scenarios. Appl. Sci. 14 (21), 9866. https://doi.org/10.3390/app14219866. based test scenario generation method for pedestrian collision avoidance system.
Wu, R.F., Zheng, X.J., Xu, Y.N., Xue, H., Lv, C., 2019. Modified driving safety field based Accid. Anal. Prev. 207, 107737. https://doi.org/10.1016/j.aap.2024.107737.
on trajectory prediction model for pedestrian–vehicle collision. Sustainability 11 Zhu, H., Iryo-Asano, M., Alhajyaseen, W.K.M., Nakamura, H., Dias, C., 2021. Interactions
(22), 6254. https://doi.org/10.3390/su11226254. between autonomous vehicles and pedestrians at unsignalized mid-block crosswalks
considering occlusions by opposing vehicles. Accid. Anal. Prev. 163, 106468.
https://doi.org/10.1016/j.aap.2021.106468.

19