Identifying factors related to pedestrian and cyclist crashes in ACT, Australia with an extended crash dataset

As vulnerable road users, pedestrians and cyclists are facing a growing number of injuries and fatalities, which has raised increasing safety concerns globally. Based on the crash records collected in the Australian Capital Territory (ACT) in Australia from 2012 to 2021, this research firstly establishes an extended crash dataset by integrating road network features, land use features, and other features. With the extended dataset, we further explore pedestrian and cyclist crashes at macro-and micro-levels. At the macro-level, random parameters negative binomial (RPNB) model is applied to evaluate the effects of Suburbs and Localities Zones (SLZs) based variables on the frequency of pedestrian and cyclist crashes. At the micro-level, binary logit model is adopted to evaluate the effects of event-based variables on the severity of pedestrian and cyclist crashes. The research findings show that multiple factors are associated with high frequency of pedestrian total crashes and fatal/injury crashes, including high population density, high percentage of urban arterial road, low on-road cycleway density, high number of traffic signals and high number of schools. Meanwhile, many factors have positive relations with high frequency of cyclist total crashes and fatal/injury crashes, including high population density, high percentage of residents cycling to work, high median household income, high percentage of households with no motor vehicle, high percentage of urban arterial road and rural road, high number of bus stops and high number of schools. Additionally, it is found that more severe pedestrian crashes occur: (i) at non-signal intersections, (ii) in suburb areas, (iii) in early morning, and (iv) on weekdays. More severe cyclist crashes are observed when the crash type is overturned or struck object/pedestrian/animal; when more than one cyclist is involved; and when crash occurs at park/green space/nature reserve areas.


Introduction
Walking and cycling are two healthy, efficient and sustainable travel modes for short distance trips.The establishment of a walkable-andcyclable community is essential to promote low-carbon lifestyles and alleviate road congestion and environmental issues.Under the influence of the COVID-19 pandemic since its outbreak, active transport modes like walking and cycling have become more and more popular.It is evident that a great amount of travellers have switched from public transport to active transport modes (Currie et al., 2021).However, along with the fast-growing population, ageing society and traffic on road, walking and cycling environments and safety issues are still severe all over the world.Pedestrians and cyclists are both considered as vulnerable road users, whereby an error that might trigger a minor incident for a vehicle occupant could lead to major consequences for pedestrians and cyclists (BITRE (Bureau of Infrastructure, Transport and Regional Economics), 2015).Evidences showed that pedestrian fatalities accounted for 13.4 % of all road fatalities in Australia.One in five people injured on Australian roads was a cyclist, and hospitalisation rates rose for cyclists but fell for other road users (AIHW (Australian Institute of Health and Welfare), 2019; BITRE (Bureau of Infrastructure, Transport and Regional Economics), 2023).Moreover, with the growing popularity of emerging transport modes on roads, pedestrians and cyclists are exposed in a more complicated and dangerous walking and cycling environments with increasing interactions to other modes, such as e-scooters, e-bicycles and autonomous vehicles.For instance, in the first few months since e-scooters were introduced to Canberra, Australia in 2020, nearly 60 residents presented to hospitals with scooter-related injuries (Riotact, 2020).
A wide field of topics on vulnerable road users' safety have been studied in recent decades, such as accident and crash analysis (Olszewski et al., 2019;Ashraf et al., 2022), exposure rating and collision risk assessment (Molino et al., 2009;Zhang et al., 2023a), and behaviour modelling and simulation (Zeng et al., 2017;Mohammed et al., 2022).Accident and crash analysis are indispensable research fields by using historical pedestrian or cyclist crash data to identify key factors affecting the frequency or severity of crashes.Identifying factors contributing to the frequency and severity of crashes is crucial to reveal the reasons behind road crashes, including roadway factors, environmental factors, vehicle factors and human factors (Chang et al., 2022;Scarano et al., 2023a).The analysis results could provide valuable insights into introducing crash countermeasures for vulnerable road users, which would be useful for designing transport infrastructure facilities and developing policies and regulations to protect vulnerable road users in various road environments (Das et al., 2019).The majority of literature studies on pedestrian or cyclist crash analysis is either at a micro-level (e.g., identifying the impact of road user demographics, traffic flow conditions, infrastructure features, and environmental factors on crashes at specific intersections or road segments) (Zahabi et al., 2011;De Rome et al., 2014;Xu et al., 2020) or at a macro-level (e.g., quantifying the spatial aggregation of crashes from a zonal perspective by considering exposure risk, socio-economic and demographic characteristics, road network and land use attributes) (Amoh-Gyimah et al., 2016;Cai et al., 2016;Gomes et al., 2017).However, limited studies have analysed pedestrian and cyclist crash data at both micro-and macro-levels, which could provide an opportunity to better understand the reasons behind vulnerable road users' accidents.Meanwhile, the impact of various factors on pedestrian and cyclist crash frequency and severity in ACT, Australia needs further investigation.Additionally, measures and recommendations should be put forward to enhance the safety of vulnerable road users.
To fill the research gap, this study establishes an extended crash dataset, which is used to identify key factors for pedestrians' and cyclists' crashes.The main contribution of this study is threefold.(1) Based on an open-source crash dataset, an extended crash dataset is established for ACT, Australia through field observation and data collection, which incorporated additional information such as road network features, land use features, and temporal features.(2) With the extended dataset, pedestrian and cyclist crashes are investigated at both macro-and microlevels to identify key factors contributing to crash frequency and severity.At the macro-level, random parameters negative binomial (RPNB) model is applied to analyse the frequency of pedestrian and cyclist crashes, and at the micro-level, binary logit model is adopted to analyse the severity of pedestrian and cyclist crashes.(3) According to the research outcomes, practical measures and suggestions are proposed for developing proper regulations and improving road infrastructure design to better protect the safety of both passengers and cyclists.
The rest of the paper is organized as follows.First, literature studies related to pedestrian and cyclist crash analysis are reviewed in Section 2. Section 3 describes the crash data and variables analysed in this study, followed by the research methodology provided in Section 4. Section 5 presents the model estimation results, followed by Section 6 with discussions on various measures and recommendations to alleviate pedestrian and cyclist crashes.Finally, the conclusion and further research are summarized in Section 7.

Literature review
Relevant studies in the literature can be categorized into two major groups as per their research purposes: identifying risk factors of pedestrian and cyclist crashes, and proposing safety measures and control strategies to alleviate such crashes and injuries.

Risk factors of pedestrian and cyclist crashes
Understanding risk factors of pedestrian and cyclist crashes is essential for reducing their injuries and improving road safety.Road user demographics are identified as significant factors relevant to crash frequency and injury severity.For example, it is found that older pedestrians were more likely to suffer severe injuries due to their greater perception and reaction times and physical vulnerability (Pour-Rouholamin and Zhou 2016;Vanlaar et al., 2016;Kemnitzer et al., 2019;Sun et al., 2019), and older cyclists were also reported to be involved in more severe and fatal injuries than younger cyclists (Kaplan et al., 2014;Behnood and Mannering 2017;Liu et al., 2020).Evidence from previous studies showed that male pedestrians and cyclists usually had a greater likelihood of suffering more severe injuries compared to their female counterparts, which could be explained by males are more likely to presented risky behaviours (e.g., traffic signal violation) (Vanlaar et al., 2016;Behnood and Mannering 2017;Scarano et al., 2023b).However, it is reported that female pedestrians and cyclists usually perceived higher risk of crashes than males (Griffin et al., 2020;Rankavat and Tiwari 2020), and females were more often involved in minor injury crashes (Das et al., 2019).Pedestrian and cyclist behaviours also potentially affect the occurrence rates of crashes and injuries on roads.Distractions and unsafe behaviours of vulnerable road users are both common risk factors contributing to fatalities and severe injuries.It is reported that pedestrians and cyclists using mobile phones, insufficient awareness of traffic conditions, violating traffic signals (or failing to give way) during walking and riding more often put them at risk of accidents (Schramm et al., 2010;Nasar and Troyer 2013;Loukaitou-Sideris et al., 2014;Hatfield and Prabhakharan 2016).Moreover, alcohol and drug use have negative impacts on human perception and decision-making, which increase the likelihood of the occurrence of fatalities and severe injuries.It is found that vulnerable road user alcohol or drugs involvement had positive associations with fatal and severe crashes (Vanlaar et al., 2016;Kemnitzer et al., 2019;Sun et al., 2019).Especially, literature results showed that alcohol consumption of cyclists sometimes was associated with other unsafe behaviours, such as low compliance of helmet use (Kaplan et al., 2014).
Road infrastructure features have significant impact on pedestrian and cyclist crash frequency and injury severity.For instance, arterial roads and intersections are found to be dangerous for both pedestrians and cyclists due to high frequency of the conflicts with motor vehicles under higher traffic volumes and higher speed limits.Zahabi et al. (2011) investigated the injury severity of pedestrians and cyclists involved in collisions with motor vehicles in Montreal, Canada, and found that pedestrian and cyclist crashes had higher risk of fatal injuries if occurred on arterial roads than these on local streets.In addition, it is reported that pedestrian crashes occurred at intersections had lower possibility of severe injury, while cyclists had higher possibility of suffering fatal injuries at intersections.Wang et al. (2015) analysed injury severity of cyclists in bicycle-vehicle crashes in Kentucky, and their results indicated that crashes occurred at uncontrolled intersections were associated with more severe cyclist injuries compared to these at stop-controlled intersections.Similarly, Samerei et al. (2021) found that the probability of fatality and serious injury of cyclists in bicycle-vehicle crashes in Australia was higher in areas without give way signs and pedestrian crossing signs.Other road infrastructure features increasing the risk of pedestrian or cyclist fatalities or serious injuries include greater motorised lanes (Wang et al., 2015;Scarano et al., 2023a), higher speed limit (Wang et al., 2015;Li et al., 2017), the presence of bus stops close to signalized intersections or greater number of bus stops in subway station service area (Heydari et al., 2017;Ashraf et al., 2022).Moreover, literature studies demonstrated that environmental conditions also had adverse effects on road safety of vulnerable road users, mainly including dark or poorly lighting roadways (Zahabi et al., 2011;Kemnitzer et al., 2019;Olszewski et al., 2019), wet or slippery road surface (Kaplan et al., 2014;Wang et al., 2015;Samerei B. Du et al. et al., 2021), and adverse weather conditions (Tay et al., 2011;Sun et al., 2019).For example, Olszewski et al. (2019) examined factors affecting fatality risk of four types of vulnerable road users in seven European countries and found that the darkness condition led to greater fatality risk for all types of vulnerable road users, but the effect appeared to be most significant for pedestrians.Kaplan et al. (2014) and Samerei et al. (2021) investigated factors affected cyclist crashes and injuries in Denmark and Australia, respectively, and their results revealed that wet or slippery road surface increased the risk of cyclist injuries and fatalities.

Safety measures and control strategies
By identifying key risk factors of pedestrian and cyclist crashes, a variety of safety measures and control strategies were proposed in previous studies to reduce their potential crash frequency and injury severity, as summarised in Table 1.First is engineering measures to create safer and more comfortable road environments for pedestrians and cyclists.Main measures and strategies include physically separating pedestrians and cyclists from motor vehicles along roads (e.g., adopting protected bicycle lanes, implementing special protections at intersections and crosswalks) (Tuckel et al., 2014;Robartes and Chen 2017;Buehler and Pucher 2021;Samerei et al., 2021), increasing bicycle lane length and improving the availability of bicycle lane (Kaplan et al., 2014;Ashraf et al., 2022), and enhancing vulnerable road users' and drivers' visibility during their trips (e.g., improving road lighting conditions and increasing sight distance) (Chen and Shen 2016;Robartes and Chen 2017;Kemnitzer et al., 2019;Sun et al., 2019;Samerei et al., 2021).Second is implementing suitable regulations or enforcement for road users to highlight traffic conflicts and reduce inappropriate behaviours.Possible measures include reducing speed limit for certain roads (Buehler and Pucher 2021), restricting mobile phone use (Nasar and Troyer 2013), implementing traffic control and traffic-calming methods (Wang et al., 2015), monitoring helmet use for cyclists (Samerei et al., 2021), and adopting improved infrastructures for pedestrians (e.g., crosswalks with activated warning lights) (Sun et al., 2019).Third is providing public education programs for vulnerable road users to increase awareness and knowledge of road users for selfprotection.Relevant recommendations include designing education programs targeting child, young and older road users (Pour-Rouholamin and Zhou 2016), launching helmet campaigns to increase helmet awareness and use (Tuckel et al., 2014;Wang et al., 2015), encouraging cyclists wearing reflective clothing and using reflectors (Chen and Shen 2016;Samerei et al., 2021), and alerting vulnerable road users and drivers avoiding alcohol and drug use (Vanlaar et al., 2016;Sun et al., 2019).

Data description
In this section, we firstly introduce the original crash data collected by the ACT Government, which is available on Open Data Portal data-ACT (Open Data Portal dataACT 2022).Based on the original crash data introduced in Subsection 3.1, additional data such as road features, land use features and other features are included for Suburbs and Localities Zone (SLZ)-based (macro-level) crash analysis in Subsection 3.2 and event-based (micro-level) crash analysis in Subsection 3.3.

Original crash data
In this study, the original crash data is obtained from Open Data Portal dataACT in ACT, Australia (Open Data Portal dataACT 2022), which provides historical records over 10 years of pedestrian and cyclist crashes from 2012 to 2021.A pedestrian or cyclist crash is defined as any crash involving at least one pedestrian or one cyclist.The crash dataset includes crash date, time, severity, crash type, number of pedestrians or cyclists involved in the crash, and reported location.The dataset records pedestrian and cyclist crashes that have been reported by the Police or the Public.There are three levels of the severity of pedestrian or cyclist crashes in the original crash data: fatal, injury and property damage only (PDO).The definition description of crash severity is provided by Open Data Portal dataACT as follows: crashes belong to fatal if there has been a fatality, crashes belong to injury if any injury has been incurred, and all the other crashes are classified into PDO.
The sample data of original crash dataset is shown in Table 2.The pedestrian crash dataset featured 11 fatal crashes, 363 injury crashes, and 189 PDO crashes, and the cyclist crash dataset featured 5 fatal crashes, 816 injury crashes, and 1067 PDO crashes.The numbers of pedestrian and cyclist crashes at different levels of severity are summarised in Fig. 1.The results show that the majority (64.48 %) of pedestrian crashes causes injury and 1.95 % of the crashes results in fatal.Unlike pedestrian crashes, the majority (56.51 %) of cyclist crashes leads to PDO. 43.22 % of cyclist crashes causes injury and only 0.26 % of the crashes results in fatal.The numbers of both pedestrian and cyclist crashes remain relatively stable before the year of 2020.Both pedestrian and cyclist crashes decreased obviously since 2020, which was probably due to the travel restrictions during the COVID-19 pandemic.

Table 1
Summary of safety measures and control strategies to reduce crash frequency and injury severity of vulnerable road users.

Category
Measure Reference

Creating safe and comfortable road environments
Separating from motor vehicles

SLZs-based data
To explore the spatial distribution of pedestrian and cyclist crashes, we utilise 136 Suburbs and Localities Zones (SLZs) in ACT based on Australian census classification in 2021.SLZs are a mesh block approximation of the officially recognised boundaries of suburbs (in cities and larger towns) and localities (outside cities and larger towns), as defined by the governments of Australia.It is worth noted that 15 SLZs are excluded from further analysis because these SLZs have no resident or a very low population in the Australian 2021 Census data.Therefore, in total, 121 SLZs are included in the analysis.Five types of features are extracted and adopted from the census data and transport related data: exposure features, demographic features, socio-economic features, road features, and other features.The census data includes details of socio-economic, demographic and travel information of the population collecting from Australian Bureau of Statistics (Australian Bureau of Statistics 2021).Transport related data includes details of road network and infrastructure of each ACT SLZ, which is collected from the ACT Open Data Portal.Similar as Hall and Tarko (2019) and Samerei et al. (2021), considering the fatality proportions of pedestrians and cyclists only account for 1.95 % and 0.26 % of all crashes, respectively, injury and fatal crashes are merged as fatal/injury (FI) crashes in the subsequent analysis.Thus, crashes are grouped into two categories based on the severity outcomes: FI and PDO crashes.Table 3 provides a summary of SLZs-based data.Number of pedestrian total crashes, number of pedestrian FI crashes, number of cyclist total crashes and number of cyclist FI crashes are set as four dependent variables.There are totally 16 independent variables.

Event-based data
To analyse the key factors related to pedestrian and cyclist crashes at

Crash frequency analysis
Poisson regression is a generalised linear approach to analyse count data.However, the underlying assumption of the Poisson distribution of variance equal to the mean is often violated in the crash data.To address the over-dispersion issue of the crash data, negative binomial (NB)  models are frequently adopted to predict crash frequency (Amoh- Gyimah et al., 2016;Tang et al., 2019).In this study, let Y i denote the observed number of pedestrian crashes or injuries, or cyclist crashes or injuries in SLZ (i = 1, 2, ..., n), X iv represent the exposure variable of SLZ i, and X ik denote the explanatory variable for SLZ i, the NB model can be specified as: wherein λ i is the parameter of Poisson model, β 0 , β v (v = 1, 2, ..., p), β k (k = p + 1,..., q) are parameters, exp(θ i ) is a gamma-distributed error term with mean and variance values equal to 1 and α, respectively.The addition of this term makes the variance differ from the mean since the variance of Y i equals to λ i + αλ 2 i .The NB model assumes that each parameter is fixed for each observation i, which may not adequately reflect unobserved differences across observations.To account for the heterogeneity in the data, random parameters negative binomial (RPNB) model is adopted, which allows the coefficients to vary across individual observations (Greene et al., 2006;Amoh-Gyimah et al., 2016): (3) wherein β iv and β ik are coefficients of vth exposure variable and kth explanatory variable for SLZ i, respectively, while δ iv and δ ik are randomly distributed terms (e.g., normally distributed terms).NLOGIT 6 is utilised to estimate NB and RPNB models.In the RPNB model, normal distribution is adopted as the distribution of random parameters, and 1000 Halton draws is used since the result of each parameter is stable with 1000 draws.A parameter will be treated as random when the standard deviation of the parameter distribution is found to be significant, otherwise it will be treated as fixed.
To ensure that highly correlated independent variables are not included in the final models, the multicollinearity issue among variables is tested and verified.The results show that the variance inflation factor (VIF) values of all variables are smaller than 6, which indicates that there is no serious collinearity among the variables.Then, Moran's 'I' is used to test for spatial autocorrelation of pedestrian and cyclist total crashes and FI crashes.Moran's 'I' statistics for pedestrian total crash and FI crash are 0.137 (Z=3.860,p < 0.001) and 0.212 (Z=4.929,p < 0.001), respectively.Moran's 'I' statistics for cyclist total crash and FI crash are 0.616 (Z=11.624,p < 0.001) and 0.601 (Z=11.172,p < 0.001), respectively.The results indicate that spatial correlations can be observed in both pedestrian and cyclist total crashes and FI crashes data.
In this study, Akaiki information criterion (AIC) is applied to evaluate the performance of the proposed models, which is calculated in Eq. ( 7).lnL is the log likelihood of the full model with predictors, and N is the number of parameters to be estimated in the model.A model with smaller value of AIC indicates a better fitting result.

Crash severity analysis
Binary logit model is frequently adopted to identify the critical factors in crash severity (Heydari et al., 2017).In this study, we set Y=0 for PDO crashes, and Y=1 for FI crashes: wherein P is the probability of FI crashes, which is equal to the number of FI crashes divided by the total number of crashes.xi is an independent variable, and β i is a model coefficient directly determining the odds ratio (OR).OR measures to what extent the outcome increases or decreases when the value of the corresponding independent variable increases by one unit.

Preliminary statistical results
Pedestrian and cyclist crashes as per the crash type is shown in Fig. 2. It can be found that the majority (83.84 %) of pedestrian crashes are "Pedestrian struck (on road)", followed by "Pedestrian struck (off road)" (14.03 %).For cyclists, 'Other -vehicle to vehicle' and 'Right angle collision' are the most common crash types, accounting for 35.01 % and 30.99 % of cyclist crashes, respectively, followed by "Side swipe" (14.94 %) and "Right turn into oncoming vehicle" (6.78 %).
Fig. 3 and Fig. 4 provide aggregated information of spatial distributions of pedestrian and cyclist total crashes and FI crashes, respectively, based on ACT SLZs.Pedestrians' total crashes and FI crashes present a relatively similar spatial distribution, with the highest value of pedestrian total crashes and FI crashes found in CITY (99 total crashes, 53 FI crashes), followed by PHILLIP (23 total crashes, 13 FI crashes), BELCONNEN (19 total crashes, 15 FI crashes) and LYNEHAM (19 total crashes, 15 FI crashes).Very few pedestrian crashes can be observed in ACT Remainder areas.Similarly, cyclist total crashes and FI crashes have a similar spatial distribution, with the highest value of cyclist total crashes and FI crashes found in CITY (204 total crashes, 78 FI crashes), followed by TURNER (160 total crashes, 65 FI crashes), BRADDON (156 total crashes, 54 FI crashes) and DICKSON (89 total crashes, 35 FI crashes).The spatial distribution of both cyclist total crashes and FI crashes appears to be more disperse than that of pedestrians.
Proportions of pedestrian and cyclist crashes as per road type is shown in Fig. 5.It is found that pedestrian crashes are more frequently observed on primary road (26.1 %) and local road (23.4 %) than on nonsignal intersection (14.4 %) and roundabout (0.9 %).On the contrary, cyclist crashes occur more frequently at non-signal intersection (33.7 %), signal intersection (19.3 %) and primary road (18.1 %).Percentages of pedestrian and cyclist crashes at different times of a day are shown in Fig. 6.The results show that pedestrian crashes occurred during inter peak (35.70 %) mostly, followed by afternoon peak (29.66 %) and morning peak (19.54 %).However, for cyclists, the crashes were more frequently found during morning peak (39.04 %), followed by afternoon peak (32.79 %) and inter peak (23.04 %).It is probably because that cycling is more often used by daily commuters during morning and afternoon peak hours.

The frequency of pedestrian and cyclist crashes
Table 5 and Table 6 present the model estimations for the frequency of pedestrian total crashes and FI crashes, respectively.Variables with significant results at 90 % significant level are presented.For both pedestrian total crashes and FI crashes, the RPNB models outperform the NB models with lower AIC values.The pedestrian total crash model (McFadden Pseudo R 2 = 0.82) fits the data better than pedestrian FI crash model (McFadden Pseudo R 2 = 0.72).Parameters with statistically significant standard deviation are regarded as random parameters.The RPNB values with bold font represent the random parameters and the others are fixed parameters.Partial effect in the RPNB model represents the effect of a 'one unit' change in an independent variable on the dependent variable.'Number of schools' is a random parameter of RPNB models (with statistically significant standard deviation) for both pedestrian total crashes and FI crashes.The results show that the frequency of pedestrian total crashes and FI crashes increases with population density, with each additional 1 thousand persons per sq.km increasing the number of pedestrian total crashes and FI crashes by 0.728 and 0.511, respectively.Higher percentage of residents driving to work results in lower frequency of pedestrian total crashes.Regarding road features, it is found that the number of pedestrian total crashes and    FI crashes increase with the percentage of urban arterial road.One explanation is that urban arterial roads usually have higher traffic volume, more intersections and crossings and more traffic signals.These features potentially increase the conflicts between pedestrians and motor vehicles, which increase the safety risks of pedestrians.Meanwhile, it is interesting that SLZs with greater on-road cycleway density have lower frequency of pedestrian total crashes and FI crashes, which can be explained by that on-road cycleway provides a safer walking environment for pedestrian by alleviating the interactions and conflicts between pedestrians and cyclists.Moreover, it is found that both pedestrian total crashes and FI crashes occur more frequently in SLZs with greater numbers of traffic signals and schools.The partial effect of the RPNB model shows that each additional traffic signal and school increase pedestrian total crashes by 0.141 and 0.453, respectively.Meanwhile, each additional traffic signal and school also increase pedestrian FI crashes by 0.113 and 0.227, respectively.The results indicate that the presence of traffic signal and school increases the safety risk of pedestrians.Suburbs with higher greater percentage of business area have higher frequency of pedestrian total crashes, with each additional 1 % of business area in SLZs increases pedestrian crashes by   0.106.One reason could be the presence of business area is associated with heterogeneous traffic volume and trip purpose and mixed land use.
In addition, number of bus stops is another factor affecting pedestrian FI crashes, with each additional bus stop increasing the number of pedestrian FI crashes by 0.027.Demographic features and socio-economic features are found with no significant influence on the frequency of pedestrian crashes.Table 7 and Table 8 present the model estimations for cyclist total crashes and FI crashes, respectively.Variables with significant results at 90 % significant level are presented.Similarly, for both cyclist total crashes and FI crashes, the RPNB models outperforms the NB models with lower AIC values.The cyclist total crash model (McFadden Pseudo R 2 = 0.94) fits the data better than cyclist FI crash model (McFadden Pseudo R 2 = 0.86).'Percentage of elderly population' and 'Percentage of rural road' are two random parameters of RPNB models (with statistically significant standard deviation) for cyclist total crashes and FI crashes.The results show that the frequency of cyclist total crashes and FI crashes decrease with the percentage of residents driving to work and walking to work, and increase with population density and the percentage of residents cycling to work.This makes sense because lower proportion of residents driving and walking to work are both associated with a lower travel frequency of cycling.SLZs with greater average household size have lower frequency of cyclist total crashes and FI crashes, which could be explained by that household with greater members may more often travel by private vehicles.Moreover, those with higher median household income have higher frequency of cyclist total crashes and FI crashes.It is found that the frequency of cyclist total crashes and FI crashes decreases with the percentage of elderly population.This is probably because that elderly population are less likely to travel by cycling compared to younger population.Higher percentage of households with no motor vehicle also increases the frequency of total crashes and FI crashes, since this promotes the household members more often travel by other modes, including walking, cycling and public transport.Taking public transport, generally, is also associated with more walking and cycling trips.In addition, high percentage of urban arterial road and rural road increases the frequency of cyclist total crashes and FI crashes, compared to other road types, including urban distributor road and urban residential road.One possible explanation is the higher traffic volume and more intersections and crossings of urban arterial road, and higher speed limit for vehicles and relatively poor protective infrastructure for cyclists of rural road might increase the safety risk of cyclists.Additionally, both cyclist total crashes and FI crashes are more often observed in SLZs with greater number of bus stops and schools.The partial effect of the RPNB model shows that each additional bus stop or school increases cyclist total crashes by 0.064 and 0.853, respectively.Meanwhile, each additional bus stop or school also increases cyclist FI crashes by 0.038 and 0.296, respectively.Moreover, it is found that the frequency of cyclist total crashes increases with the number of traffic signals.However, the impact of the number of traffic signals on cyclist FI crashes is not significant.

The severity of pedestrian and cyclist crashes
The influencing factors of crash severity of pedestrians are analysed, and the results are as shown in Table 9. Crash type is found to be significantly related to crash severity of pedestrians, and the odds of the FI crashes in 'Other types' (including 'Right angle collision', 'Right turn into oncoming vehicle', 'Collision with parked vehicle', 'Opposite direction side swipe' and 'Other − Vehicle to Vehicle') are 12.904 times of that in 'Struck pedestrian (On Road)'.The results reveal that the odds of FI crashes of pedestrians at signal intersection are lower than that on non-signal intersection, and higher than that on roundabout, primary road, local road, and other road types.In the meanwhile, the severity of pedestrian crashes is lower when light rail station locating within 150 m of the crash location (OR=0.078,p = 0.063).Compared to pedestrian crashes in city areas, these reported in suburb areas tend to be more serious (OR=1.691,p = 0.064), which highlights the importance of speed limit regulation and adequate walking infrastructure for pedestrians.Pedestrians are more likely to encounter FI crashes in early morning (0am-6:00am) than other time periods, especially higher than morning peak (6:00am-9:30am) (OR=0.312,p = 0.036).This is probably owing to low lighting conditions of roadway and low attention of drivers and vulnerable road users during the early morning period.Moreover, day of the week appears to be another important factor affecting crash severity of pedestrians.Compared to weekdays, FI crashes of pedestrians is less likely to occur on weekend and holiday (OR=0.575,p = 0.019).Other factors, including involved pedestrian(s), if close to bus stop and traffic signal, are not found significantly affecting the crash severity of pedestrians.
The influencing factors of crash severity of cyclists are analysed, and the results are as shown in Table 10.Similar to the results of crash severity of pedestrians, crash type is a significant factor affecting crash severity of cyclists.'Right turn into oncoming vehicle', 'Collision with parked vehicle', 'Overturned' and 'Struck object, pedestrian or animal' crash types are more likely to result in FI crashes of cyclists, compared to 'Right angle collision' crash types.It is found that 'Overturned' and 'Struck object/pedestrian/animal' are two crash types with most serious crash severity.In 32 'Overturned' cases, 1 fatality and 26 injuries were recorded, and FI crashes account for 84.4 % of all crashes.In 45 'Struck object/pedestrian/animal' crashes (the majority is 'Struck pedestrian'), 1 fatality and 38 injuries were recorded, and the FI crashes account for 86.7 % of all cases.As a reference, in 585 'Right angle collision' crashes, 1 fatality and 245 injuries were recorded, and FI crashes account for 42.1 % of all crashes.The crash might become more severe when more than one cyclist is involved in the crash than the case with single cyclist involved (OR=2.981,p = 0.015).Although greater number of bus stops in SLZs leads to higher frequency of cyclist crashes and FI crashes, the presence of bus stop within 50 m can mitigate the crash severity of cyclists.This is probably associate to the lower speed of both cyclists and vehicles (including buses) near bus stops.In the meanwhile, the results show that land use is significant to crash severity of cyclists.For example, it is found that the occurrence odds of FI crashes of cyclists at park, green space and nature reserve areas are 1.722 times of that of cyclists at commercial and business areas, and the odds at other land use are 1.981 times of that at commercial and business areas.Other factors, including road type, if close to traffic signal and light rail station, if occurs in city or suburb areas, are not found significantly affecting the crash severity of cyclists.

Discussion
Vision Zero was adopted since last century aiming to eventually achieve zero road death and serious injury on the road transport network (Shi et al., 2023).To achieve the target of Vision Zero, the ACT Government has implemented a variety of road safety measures for saving lives and reducing injuries (ACT Government 2020).How to effectively reduce the actual and perceived risks of injury and death from walking and cycling remains a challenge (Buehler and Pucher 2021).Based on the research findings in this project, multiple suggestions are proposed in this section to help minimise the potential safety risk while create a safe road environment for vulnerable road users.
The research findings indicate that a high percentage of urban arterial road increases the frequency of pedestrian total crashes, cyclist total crashes and cyclist FI crashes, which can be a major factor related to crashes (Dumbaugh and Li 2010;Ma et al., 2010).One possible reason is that high traffic volumes of vulnerable road users and vehicles are more frequently observed on urban arterial roads.Moreover, urban arterial roads usually have more lanes, intersections, crossings, traffic signals, and are closer to business area, which also result in higher safety risks of crossing behaviours.One way to alleviate this issue is creating a safe and comfort road space for different types of road users, such as setting overpasses and underground passages for pedestrians and cyclists on busy roads to alleviate the interactions between them and motor vehicles.Moreover, it is found that a high proportion of rural roads also increases the frequency of cyclist total crashes and FI crashes.This is probably because that motor vehicles and cyclists have higher speed on rural roads, and generally rural roads lack cycleway provisions or without on-road cycleways for cyclists (Carter and Council 2006;Chang et al., 2022;Rella Riccardi et al., 2022).Therefore, various countermeasures, such as setting speed limit for motor vehicles and cyclists, providing more sufficient riding space, and implementing warning signing for cyclists, could be taken into account to reduce the potential safety risk.
It is evident that pedestrian and cyclist safety at intersections and crossroads is one of the most critical issues (Hu et al., 2018).It is found that the frequencies of pedestrian total crashes, pedestrian FI crashes and cyclist total crashes increase with the number of traffic signals.Intersections and crossroads are dangerous due to the high risk of the interactions between pedestrians (and cyclists) and motor vehicles.Meanwhile, traffic signal violation behaviours of motor vehicle drivers and vulnerable road users also contribute to a large number of pedestrian and cyclist crashes (Zhang et al., 2016).One literature study reported that a number of characteristics at traffic signals would lead to more pedestrian crashes, including higher volumes of pedestrian and motor vehicle traffic, longer crossing distances, more crosswalks and greater intersection density (Islam et al., 2022).Therefore, it is significant to provide clear signals and protective facilities for pedestrians and cyclists and avoid the traffic signal violation behaviours, such as setting refuge island for long-distance crossing and enhancing public educations on road regulations.
Our findings also reveal that the frequency of pedestrian FI crashes, cyclist total crashes and cyclist FI crashes increase significantly with the number of bus stops.Pedestrians and cyclists may suffer from severe safety risks near bus stops.Recent studies indicated that bus stops were highly risky facilities with relatively higher number of pedestrian fatalities (Lakhotia et al., 2020), and there was spatially correlations between pedestrian crashes and bus stop locations (Ulak et al., 2021).In the meanwhile, passengers boarding or alighting buses may encounter unavoidable conflicts with passing cyclists, which could lead to delayed passing and potential collision risk for both bus passengers and cyclists (Zhang et al., 2023a).Thus, effective strategies are needed to minimise the crashes occurring near bus stops.For example, bus bunching should be avoided to alleviate boarding or alighting of a large volume of passengers simultaneously.Besides, layout design of bus stops should provide sufficient space and clear view for bus passengers, pedestrians and cyclists, to make sure they are able to notice potential risk and present evasive behaviours in advance.
It is evident that pedestrian FI crashes, cyclist total crashes and cyclist FI crashes are more likely to be observed in suburbs with a greater number of schools.This phenomenon could be resulted from the fact that there are relatively a large proportion of school-age child pedestrians and cyclists around schools.Moreover, the traffic conditions near schools could be more complicated, which may pose adverse impacts on passing pedestrians and cyclists.The results from an early study showed that there were close relationships between the school (and the neighbourhood attributes) and pedestrian crashes near schools (Clifton and Kreamer-Fults 2007).Another recent study found that proximity to school increased the frequency of pedestrian injuries at nearby intersections (Heydari et al., 2020).To deal with this issue, it is crucial to implement school-zone speed limits strictly to reduce injury severity of pedestrian and cyclist crashes.In addition, main infrastructures, including footpaths, cycleways, and shared paths around schools, could be equipped with dedicated fence systems to provide further protection to school-age pedestrians and cyclists.Additionally, the frequency of pedestrian total crashes decreases significantly with the increase of on-road cycleway density.One important reason is that there are less interactions and conflicts between pedestrians and cyclists if cyclists ride on on-road cycleways more often.Recently, the coexistence of pedestrians and cyclists in shared space raises many concerns on road safety (Zhang et al., 2023b).The literature studies highlighted that collisions between cyclists and pedestrians could result in serious injury (Chong et al., 2010), and the most severe injuries of pedestrians more likely stemmed from the secondary impacts of collisions with cyclists, such as falling and hitting the head on the ground (Graw and König 2002).Besides, a study investigated cyclist crashes in different riding environments in ACT, Australia, and found that fewer cyclists were injured in on-road cycleways than in other cycling environments, and a high proportion of injuries were incurred on shared paths (De Rome et al., 2014).Our results confirm that cyclist struck pedestrian is one of the most dangerous crash types for cyclist, which in line with the literature findings.Proper strategies can be taken to alleviate the corresponding safety risk.For instance, suburbs with high traffic volume are recommended to utilise separated road facilities to guarantee the safety and increase the comfort level of vulnerable road users.Suitable regulations like space sharing rules should be implemented to minimise the illegal riding behaviours.
Finally, most severe pedestrian crashes are found to occur in suburb areas and during early morning, and most severe cyclist crash types are overturned and struck object/pedestrian/animal, which highlights the importance of improving road conditions for pedestrians and cyclists.It is crucial to create sufficient infrastructures and adequate facilities to provide positive perceptions of safety and comfort during walking and cycling.For example, transport planners and engineers could alleviate pedestrian and cyclist injury risks by improving street lighting condition, ensuring road surface smoothness, and removing obstacles from walking and cycling paths.

Conclusion
This study explored key factors correlated with pedestrian and cyclist crashes in ACT, Australia.Pedestrian and cyclist crash frequency and severity were analysed using SLZs-based variables and event-based variables, respectively.The research findings show that most pedestrian crashes cause injury while the majority of cyclist crashes lead to property damage only.Many factors are found to be associated with high frequency of pedestrian total crashes and FI crashes, including high population density, high percentage of urban arterial road, low on-road cycleway density, high number of traffic signals and high number of schools.Meanwhile, multiple factors show positive relations with high frequency of cyclist total crashes and FI crashes, including high population density, high percentage of residents cycling to work, high median household income, high percentage of households with no motor vehicle, high percentage of urban arterial road and rural road, high number of bus stops, and high number of schools.In addition, it is found that more severe pedestrian crashes are more often occurred at nonsignal intersections, in suburb areas, during early morning, on weekdays, or without light rail station locating within 150 m.More severe cyclist crashes are more likely to observed when crash type is overturned or struck object/pedestrian/animal, when more than one cyclist involved, without any bus stop within 50 m, or at park/green space/ nature reserve areas.
This study conducted pedestrian and cyclist crash analysis at both macro-and micro-levels to identify key factors related to crash frequency and severity.RPNB models were adopted to analyse crash frequency affected by SLZs-related features (e.g., exposure features, demographic features, and socio-economic features) at the macro-level, while binary logit models were used to evaluate crash severity influenced by event-related features (e.g., road type, land use, and time of day) at the micro-level.The combined research findings provide suggestions for urban and rural planning in macro-perspective and infrastructure design and road regulation development in microperspective to improve the safety of pedestrians and cyclists.The implications are not only applicable to ACT region but also adding insights to other cities especially for those planning new residential and commercial development zones, where designs must consider all modern road users.However, this study has limitations.The interpretation of the implemented models relies on the availability of relevant variables and attributes of crash data.Some variables, such as traffic volume, speed limit and lighting condition at different road segments, were not available in this study, which however may affect the analysis results.In the future work, it is recommended to collect and analyse more broad data for further investigation.For example, it would be helpful to collect road user's demographic and behavioural information, such as gender, age group, and distracting behaviours.Beyond the existing crash dataset, it is a black spot for identification and analysis of underreported crashes or near-crashes of vulnerable road users under various road environments.

Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Fig. 3 .
Fig. 3. Spatial distributions of pedestrian total crashes and FI crashes in ACT.

Fig. 4 .
Fig. 4. Spatial distribution of cyclist total crashes and FI crashes in ACT.
B.Du et al.

Fig. 5 .
Fig. 5. Proportions of pedestrian and cyclist crashes as per road type.

Fig. 6 .
Fig. 6.Pedestrian and cyclist crashes at different times of a day.

Table 2
Sample data of original crash dataset.-level, an extended crash dataset is established based on each crash record.With the location information (i.e., latitude and longitude) from the original crash data, road network features and land use features are extracted from OpenStreetMap by implementing an Overpass API query with Python.After extracting relevant features, both online cross-check using Google Map and onsite inspection at sample locations (20 % of the total crash locations) in the ACT are conducted to validate the extended crash dataset, with an accuracy of 80 % and 90 %, respectively, of the road network features and land use features.These steps help to identify any discrepancies and exclude error data for further analysis.
Fig. 1.The numbers of crashes at different levels of severity over years.B.Du et al.micro

Table 3
Summary of statistical data of SLZs-based variables.

Table 4
Summary of statistical data of event-based variables.

Table 5
Model estimation for the frequency of pedestrian total crashes.
B.Du et al.

Table 6
Model estimation for the frequency of pedestrian FI crashes.

Table 7
Model estimation for the frequency of cyclist total crashes.

Table 8
Model estimation for the frequency of cyclist FI crashes.

Table 9
Model estimation for the severity of pedestrian crashes.

Table 10
Model estimation for the severity of cyclist crashes.