Download as pdf or txt
Download as pdf or txt
You are on page 1of 15

Accident Analysis and Prevention 142 (2020) 105538

Contents lists available at ScienceDirect

Accident Analysis and Prevention


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

Simulating Automated Emergency Braking with and without Torricelli T


Vacuum Emergency Braking for cyclists: Effect of brake deceleration and
sensor field-of-view on accidents, injuries and fatalities
Hanna Jeppsson*, Nils Lubbe
Autoliv Research, Wallentinsvägen 22, 44783 Vårgårda, Sweden

ARTICLE INFO ABSTRACT

Keywords: This study estimates how many additional cyclist accidents, injuries or fatalities are avoided or mitigated by
Autonomous Emergency Braking adding a system which increases braking levels, the Torricelli Vacuum Emergency Brake (VEB), to a state-of-the-
Bicycle art Automated Emergency Braking (AEB) system. To obtain a realistic state-of-the-art AEB system, the AEB
Euro NCAP parameter settings were defined to fulfil but not exceed the performance necessary to achieve a full score in the
Injury risk curves
European New Car Assessment Program (Euro NCAP). The systems are simulated in a simple but realistic si-
Simulation
mulation model in MATLAB with varying brake deceleration and sensor field-of-view (FoV).
This study utilised data from the German In-Depth Accident Study (GIDAS), released in January 2019, and the
related Pre-Crash Matrix (PCM), released in February 2019. Cyclist Injury Risk Curves were created from 2,662
GIDAS accidents involving a passenger car and a cyclist. The sample of cyclist accidents from the GIDAS-PCM
database used in the final simulations comprised 1,340 collisions between the front of a passenger car and a
cyclist. Both data samples were weighted to be representative of Germany as a whole.
Adding the VEB was found to avoid over 20% more accidents than the AEB alone. Although increasing the
FoV from 75° to 180° for the AEB system increases its accident avoidance rate to a level comparable to the VEB,
the VEB remains about 8–20% more effective in reducing fatalities and injuries, and thus offers greater safety
benefits than simply increasing AEB FoV.
While the initial accidents in the representative simulation sample are fairly evenly distributed over the
vehicle front, the remaining accidents (those that cannot be prevented by AEB or VEB) are more concentrated at
the vehicle corners and are further characterized by high cyclist speeds. High cyclist speeds and impact to the
vehicle corners potentially increase the relative frequency of head impacts to the stiff A-pillars. We therefore
recommend that, for passenger cars, VEB and other advanced AEB systems should be combined with in-crash
protection, especially in the A-pillar area, to best protect cyclists from injury.

1. Introduction statistics for France indicate that 72% of cyclist fatalities occur in ac-
cidents with one other traffic participant, 66% of which were passenger
Cycling as a mode of transport contributes to sustainable mobility cars (ONISR, 2017), while in Germany these figures rise to 78% and
and, being a physical activity, to better health. However, the risk of 75% (Statistisches Bundesamt (Destatis), 2018a). In Sweden, although
injury through cycling in traffic, with consequences as severe as death, 48% of cyclist deaths do not involve other traffic participants, of those
is a major drawback and for many people a barrier to adopting this fatalities involving motorized vehicles, 58% involve passenger cars
mode of transport (ITF, 2019). Cyclist fatalities are far too common (Trafikanalys, 2018). While these numbers are alarming, they may in
around the world: an estimated 69,000 cyclist lives are lost every year fact underestimate the true epidemic, as accidents involving cyclists are
and another 11 million cyclists are injured (James et al., 2017; Roth the most underreported of all road crashes (Elvik & Mysen, 1999; Held,
et al., 2018). In the Netherlands, a country where cycling is popular, 2016; Isaksson-Hellman & Werneke, 2017). For instance, in Sweden,
cyclists account for as much as 30% of all road traffic fatalities (OECD/ 90% of cyclist accidents are not reported (Held, 2016).
ITF, 2018). Today’s transport system can be improved to better protect cyclists
Many cyclist fatalities occur in crashes with passenger cars. Recent through changes to infrastructure, through increasing the use of cyclist


Corresponding author.
E-mail address: Hanna.Jeppsson@autoliv.com (H. Jeppsson).

https://doi.org/10.1016/j.aap.2020.105538
Received 8 November 2019; Received in revised form 31 March 2020; Accepted 1 April 2020
Available online 26 May 2020
0001-4575/ © 2020 Elsevier Ltd. All rights reserved.
H. Jeppsson and N. Lubbe Accident Analysis and Prevention 142 (2020) 105538

protective equipment, and through more cyclist-friendly vehicle design. et al., 2016). Euro NCAP performs two different AEB test scenarios
Suggested improvements to infrastructure include separate, well- involving cyclists, one where the cyclist is crossing the path from the
maintained cycle paths and safe road crossings (Miljö-och en- near-side of the vehicle and one where the cyclist is travelling in the
ergidepartementet, 2018; Wegman et al., 2012). Cyclist protection can same direction as the vehicle, this latter in two configurations (Euro
mitigate injury outcome substantially: the most common form, bicycle NCAP, 2019a). Further scenarios are planned from 2020, possibly to
helmets, are one part of the Swedish path to Vision Zero include crossing from far-side (Euro NCAP, 2019b). The set-up com-
(Näringsdepartementet, 2009), and a systematic review determined prises a bicycle mounted on a mobile platform and a special dummy
that the odds ratio of a cyclist sustaining serious head injuries in an with movable legs which imitates a cyclist’s motion. To achieve a full
accident or fall when wearing a helmet compared to not wearing one test score in the cyclist crossing scenario, Euro NCAP requires the ve-
was 0.3 (Olivier and Creighton, 2017). Another systematic review by hicle to come to a complete stop when travelling at speeds of up to
Esmaeilikia et al. (2019) found that wearing a bicycle helmet was not 40 km/h, and vehicle speed to be reduced by at least 20 km/h when
substantially associated with engaging in risky behaviour. Hence, travelling at speeds of between 40 km/h and 60 km/h (Euro NCAP,
helmet wearing does not increase the number of accidents, but reduces 2017).
their severity, and is therefore widely promoted. Other personal pro- Improving the performance of AEBs to prevent more crashes can
tection equipment offers benefits as well, such as high-visibility gear increase the frequency of unnecessary brake activations. If this is done
(Wood et al., 2009) and shoulder pads (Stigson and Fahlstedt, 2016). people might choose not to have AEB, either by switching it off, not
Some of these solutions have not been widely adopted, possibly due to buying it, or buying a different one which is less intervening and thus less
inconvenience, cost or lack of awareness. Finally, cyclist-friendly ve- effective. One solution to this problem is the Torricelli Vacuum
hicle design can prevent or mitigate cyclist accidents using techniques Emergency Brake (VEB), which increases braking force without in-
similar to those used to protect pedestrians, for example by reducing creasing the frequency of unnecessary brake activations. When activated,
collision speed, by softening the impact with the car front end, and by the VEB deploys a vacuum pad onto the road surface. The pad is con-
absorbing energy with sufficient deformation space and within human nected rigidly to the vehicle frame and flexibly to a vacuum tank in-
tolerance (Haddon, 1970; Hu and Klinich, 2012). Cyclist-friendly de- stalled on the vehicle underbody via a hose. Once deployed, the vacuum
signs can include deployable hood lifters that increase the deformation pad exerts a pull force against the road surface which transfers through
space before hard structures in the engine bay are reached (Fredriksson the rigid link to the vehicle, which, in turn, increases the normal force in
et al., 2001; Ames and Martin, 2015). Other complementary solutions the tire–road contact areas. The increased normal force leads to an in-
are external airbags to soften impact with the stiff windscreen base and creased braking force, given a constant road friction coefficient and ve-
A-pillars (Fredriksson et al., 2015) which otherwise can result in high hicle mass, while the duration of the increased force is dependent on the
head accelerations and head injury risk in car-to-cyclist accidents size of the vacuum tank and the sealing of the vacuum plate. In this way,
(Katsuhara et al., 2014). namely through braking harder than a reference design but not acti-
Collision prevention technologies also have great potential for vating earlier, it has been found that for a given unnecessary activation
cyclist protection (Silla et al., 2017). Automated Emergency Braking rate up to 22% more pedestrian injuries and 15% more pedestrian
(AEB) for cyclists is particularly promising: when an imminent collision fatalities can be prevented (Jeppsson et al., 2018).
with a cyclist is detected by the vehicle sensors, the system is designed Although research has begun to quantify the safety benefits of the
to apply the brakes automatically, reducing vehicle speed and ideally VEB for pedestrians, its potential benefit for cyclists remains unknown.
bringing the vehicle to a safe stop. Such AEB systems include sensors, Using virtual simulations of German pre-crash data, the aim of this
brakes (and their stopping performance), and an activation logic. Rosén study therefore is to determine the benefits of VEB for cyclists in terms
(2013) estimated that AEB can reduce severe injuries in cyclist acci- of accidents, injuries and fatalities avoided by comparing the perfor-
dents by 5%–67%, and fatalities by 6%–84%. This wide range of pos- mance of the VEB with an AEB. The study comprises four steps:
sible outcomes depends on the system settings: the best performance First, we define a simple but realistic AEB model. This model si-
was obtained when AEB was activated 1.5 s prior to the collision, when mulates the three core elements of an AEB system: sensor coverage,
there were no operating restrictions, when the sensor field-of-view activation logic, and brake performance. The aim is to determine the
(FoV) was wide at 90°, and with a brake deceleration of 10 m/s2; minimum value of these three elements needed to achieve a full score in
conversely, the poorest performance was obtained when the system was Euro NCAP; such systems are considered as state-of-the-art, balancing
activated only 0.5 s before collision, with restrictions of not operating in what is technically possible with what is economically feasible.
darkness or above 60 km/h, with a limited sensor FoV of 40°, and a Secondly, we add the VEB to the AEB, referred to as VEB. In this
brake deceleration of 5 m/s2. case, while deceleration is greater, brake activation occurs at the same
Sensor FoV is of particular importance to AEB performance. For time as for the state-of-the-art system and thus the unnecessary acti-
pedestrian protection, increasing a 20° FoV to 40° or to 180° increases vation rate is the same as for AEB only, referred to as AEB. We estimate
the rate of avoided fatalities from 30% to 40% and 44%, respectively how many additional accidents, injuries and fatalities VEB will avoid or
(Rosén et al., 2010). mitigate by calculating and applying risk curves for each accident.
For cyclists, FoV is even more important as their greater speed can Thirdly, we increase the FoV to 180° in order to predict the ability of
place them even further away from the car at the critical time. Using more advanced AEB systems to prevent cyclist accidents, and compare
German in-depth data, Ranjbar (2014) investigated cyclist positions 2 s this to the VEB. As an additional comparison, a VEB system with a FoV
prior to crash and noted that a FoV of 50° can cover 50% of severe of 180° is also evaluated.
accidents, while a FoV of more than 100° is necessary to cover 90%. In Finally, we characterize the residual accidents, those which are not
particular, turning scenarios were noted to require a large FoV. Similar prevented, based on accident scenario, impact location on vehicle and
findings are reported by Zhao et al. (2019) based on AEB simulations of driving speed in order to highlight developments that can be made to
a collection of 40 cases from Japan. AEB was predicted to be able to protect cyclists. We also compare the Euro NCAP AEB protocol re-
avoid 15% of cyclist collisions with a 50° FoV, which increased to 50% garding cyclists and investigate how it differs from our dataset in terms
with a 90° FoV. Notably, even with an ideal AEB system with a 360° of test scenarios and vehicle and cyclist travelling speeds.
FoV, not all accidents can be prevented: impacts remained in 20% of the
simulated cases. 2. Method
The European New Car Assessment Program, Euro NCAP, aims to
provide consumers with an independent safety rating for new vehicles Mathematical models of AEB and VEB systems were incorporated
and to influence the development of new safety systems (van Ratingen into simulations of reconstructed accidents to quantify how such

2
H. Jeppsson and N. Lubbe Accident Analysis and Prevention 142 (2020) 105538

systems influence these accidents. Their most important influence is 2.2. Euro NCAP scenario
known to be the reduction of impact speed (Zhao et al., 2019). Acci-
dents may be completely prevented or their severity mitigated. Miti- To obtain a realistic state-of-the-art AEB system, the parameters are
gation is quantified by linking impact speed to injury outcome with pre- set to give the AEB a full score in Euro NCAP, but not to give a per-
defined injury risk curves; their construction is an important part of this formance which exceeds this. The minimum setting of sensor coverage
study. and activation logic was thus based on the most challenging scenario
The method section is structured as follows. First, the simulation from the Euro NCAP VRU protocol, which is where the cyclist is
model, which calculates the effect of AEB and VEB on the accidents, is crossing the vehicle’s path (Euro NCAP, 2019a). The values taken from
introduced. Second, the tuning of the AEB parameters for Euro NCAP the VRU protocol below were used as a basis to set the system para-
scenarios is described. Third, the data selection process to obtain the meters:
sample of reconstructed accidents is defined. Fourth, the weighting
procedure to make the data sample representative of Germany as a • Friction coefficient, μ = 0.9
whole is given. Fifth, the weighted data sample is described. Sixth, the • Vehicle velocity, v = 11.1 m/s and 5.56 m/s
vehicle (40 km/h and
construction of the injury risk curves is explained and lastly, seventh, 20 km/h)
system effectiveness is defined. • Bicycle velocity, v = 4.16 m/s (15 km/h).
cyclist

2.1. Simulation model The Ramp‐up time, the time that it takes before full brake pressure
is reached, tramp, was set to 0.15 s as for a state-of-the-art AEB, following
A simulation script in MATLAB was used to simulate an AEB and Edwards et al (2014).
VEB systems. The simulation model was created by Rosén (2013) and Eq. (1) gives the deceleration a, which is used in Eq. (2) to de-
adapted for this study using newer data with an increased number of termine the distance it takes for the vehicle to come to a complete stop,
accidents, updated injury risk curves, and updated system parameters. dvehicle, (the derivation of this formula is given in Appendix C). This
A large number of parameters define the systems (Appendix A); how- brake distance can be calculated when integrating constant jerk in the
ever, only two have been varied in this study: sensor FoV to simulate ramp-up phase and constant deceleration in the braking phase as-
more advanced sensing systems, and brake deceleration to simulate the suming zero acceleration and constant speed at brake initiation. Eq. (3)
function of the VEB. The brake deceleration, a, was calculated for AEB calculates TTC at brake initiation when the AEB needs to start braking,
systems as μ*g, where μ is the friction coefficient between the vehicle TTCaeb. In Eq. (4) the distance the cyclist travels, dcyclist, during TTCaeb is
tires and the road surface and g is the gravitational constant, 9.81 m/s². calculated, which gives the distance of the cyclist from the collision
For VEB, a was calculated as μ*1.8*g which corresponds to a VEB of size point.
0.6 × 0.6 m.
a = µ *g (1)
The AEB and VEB system activates and intervenes by braking if the
following criteria are met (Fig. 1): vvehicle 2 vvehicle tramp atramp2
d vehicle = +
2a 2 24 (2)
a) at least half of the cyclist is visible (defined by the center of gravity
(CoG) being in the sensor FoV and range) as the AEB cannot activate TTCaeb =
d vehicle
if the cyclist is not visible; vvehicle (3)
b) the cyclist is inside the trigger width, to limit unnecessary activa-
tions for cyclists far away from the driving path; dcyclist = TTCaeb*vcyclist (4)
c) predicted Time To Collision (TTC) is less than TTCaeb, i.e. the ear- TTC at trigger decision to come to a complete stop, TTCaeb, at the
liest time that the AEB is allowed to activate (a static activation maximum speed required by Euro NCAP, 40 km/h, is calculated using
threshold, in order to limit unnecessary activations); Eqs. (1–3) to be 0.7 s. The distance that the bicycle travels during this
d) predicted TTC is less than TTCstop, i.e. the time at which an un- time gives a lateral offset to the calculated collision point which ac-
braked car needs to start braking in order to avoid the accident (in cording to Eq. (4) is 2.9 m.
order to delay activation until the latest time when crash avoidance In the Euro NCAP scenario, the impact point between the partici-
is still feasible), see explanation in Appendix B; pants is defined as being the middle of the vehicle’s front and 77 mm in
e) the vehicle is predicted to collide with the cyclist, i.e. at least one of front of the bicycle’s CoG (Euro NCAP, 2019a). Recall that the trigger
the bicycle’s four corners needs to be within the predicted car path width in the simulation model is defined from the predicted driving
(depicted as four individual predicted impact points, one for each path to the front corner of the bicycle (Fig. 1). To be able to calculate
corner). the trigger width, dcyclist needs to be adjusted by vehicle width and bi-
cycle dimensions, Eq. (5), where wcar is 1.73 m (calculated as the
average width of the vehicles in the data sample) and dcyclistCoGtoFront is
0.96 m (the average distance of the CoG to the front of the bicycles in
the same dataset).
wcar
Trigger width = dcyclist dcyclistCoGtoFront + 0.077
2 (5)

The TTCaeb decision occurs at a trigger width of 1.2 m, Eq. (5).


To determine sensor FoV, it is not the highest but the lowest car
driving speed which is the most challenging. For lower driving speeds
dvehicle will decrease which will make the angle to bicycle greater and
thus give a wider FoV. Having calculated the distances that both the
vehicle, dvehicle, Eq. (2), and bicycle, dcyclist, Eq. (4), travel during TTCaeb,
the FoV can be calculated by using the inverse tangent. Using the lowest
vehicle speed tested in Euro NCAP VRU will give the widest FoV
needed; using 20 km/h therefore gives the widest necessary FoV of 75°.
Fig. 1. Decision algorithm for the simulation model. This AEB system achieves the highest score in the Euro NCAP

3
H. Jeppsson and N. Lubbe Accident Analysis and Prevention 142 (2020) 105538

scenarios and is referred to as Baseline 75°. Three additional variations 2.5. Data characterization
were evaluated: this baseline system with a larger FoV, called Baseline
180°; the baseline system with VEB added, called VEB 75°, and the The (weighted) sample is characterised by impact speed of the ve-
baseline system with both the VEB and a larger FoV, called VEB 180°. hicle and cyclist, location of first impact on the vehicle front, and ac-
cident scenario. Vehicle impact speed has a mean of 5.9 m/s and
2.3. Simulation data standard deviation of 4.5 m/s. Cyclist impact speed has a mean of
4.4 m/s and standard deviation of 1.6 m/s. The impact points on the
The German In-Depth Accident Study (GIDAS) is a database that vehicle fronts are distributed over the complete vehicle front with the
contains detailed information about traffic accidents occurring in and corners being impacted at marginally higher rates. The accident sce-
around the German cities Dresden and Hanover. Special investigation narios were categorised into twelve groups, adopted from Ranjbar
teams began to collect information about accidents in these areas in (2014): first the vehicle is categorised into one of three groups de-
1999 (Erbsmehl, 2009); over 30,000 accidents have been analyzed pending on its movement: going straight, turning left or turning right;
since then, with almost 2,500 variables collected per case. then the cyclist is categorised into one of four groups depending on
The GIDAS Pre-Crash Matrix (PCM) contains two-participant acci- their movement: same direction, opposite direction, left or right prior to
dents from the GIDAS database which have been reconstructed to the collision. The movement of both the vehicle and cyclist are col-
provide detailed information of the trajectories of both participants lected at the same time as TTCaeb, i.e. 0.7 s before the accident. The top
involved. Data for the final seconds before the crash is reconstructed three scenarios representing 65% of the accidents are: cyclist crossing
with a sample frequency of 100 Hz (Erbsmehl, 2009). In GIDAS-PCM, the path of the vehicle from the right while the car is going straight,
all passenger cars are visualized as rectangles with cut corners in the cyclist crossing the path of the vehicle from left while the car is going
front of the vehicle and cyclists are visualized as diamonds with the straight, and cyclist crossing path of the vehicle from right while the car
widest part being the handlebar of the bicycle (VUFO, 2019). is turning right.
This study uses GIDAS and GIDAS-PCM data released in January In the results section, the distribution of the characteristics of the
and February 2019, at which time GIDAS contained 8,600 accidents sample are depicted together with those of the remaining accidents.
involving injured cyclists. Of these accidents, 2,731 were reconstructed
in the GIDAS-PCM database. Keeping only cases where the cyclist was 2.6. Injury risk curves
hit by a passenger car (defined by UN-ECE class M1) reduced the count
to 1,962 and restricting these to frontal impacts reduced the count Cyclist Injury Risk Curves (IRC) evaluate the risk of an injury to a
further to 1,904. Newer vehicles are designed to be less likely to injure cyclist in an accident as a function of vehicle impact speed. To calculate
VRUs (European Commission, 2016); therefore, this study focuses only these IRCs, GIDAS data released in January 2019 were used. Out of
on newer vehicles. Removing cases with vehicle registration year 1997 8,600 accidents involving injured cyclists, 5,039 remained when the
or earlier decreases the number to 1,394. Lastly, cases with unknown count was restricted to cyclists impacting the front of a vehicle. Of
Maximum Abbreviated Injury Scale (MAIS) level were removed, re- these, 4,162 were struck by a passenger car. Restricting the count to
sulting in a sample of 1,340 accidents. Three of the vehicles were coded newer vehicles by removing cases with registration year 1997 or earlier
as being equipped with a cyclist- or pedestrian-AEB system. However, decreased the number to 2,907. Finally, cases with unknown MAIS level
no braking occurred prior to the accident and hence we assume that the were removed, resulting in 2,662 accidents. The risk curves were cre-
AEB was not activated; these accidents were therefore included in the ated for three levels of injury severity, fatal, at least serious (MAIS3+),
sample. and at least moderate (MAIS2+). The GIDAS data was weighted to
Germany (Table 2), using the official definition of injury levels, fol-
2.4. Weighting lowing the same approach as for the GIDAS-PCM weighting described
earlier. Vehicle speed had a mean of 4.6 m/s with a standard deviation
Since GIDAS has a disproportionate number of severe and fatal ac- of 3.3 m/s; for cyclist speed these were 3.9 m/s and 1.6 m/s, respec-
cidents, the data must be weighted to be representative for Germany as tively. Cyclist age ranged from 3 to 88 years with a mean of 44 years.
a whole. Using data from German national road traffic statistics for Male cyclists accounted for 54%. Only 13% of the cyclists wore some
2017 (Statistisches Bundesamt (Destatis), 2018b), weight factors can be kind of helmet.
calculated in two steps. First, relative weight factors are calculated for Logistic regression was performed using the glm() function in R,
all passenger car to bicycle accidents in the GIDAS-PCM (1,962 cases) to calculating the injury risk in the form of Eq. (6)
represent the distributions of fatal, severe and slight injuries in Ger- risk (v ) = e b0 + b1 v /(1 + e b0 + b1 v ) (6)
many. Secondly, these factors are normalized for the sample data so
that unweighted and weighted sample sizes (1,340 cases) are equal In Eq. (6), b0 and b1 are the regression coefficients and v is the ve-
(Sander and Lubbe, 2018). The final, weighted sample consists of 3.4 hicle impact speed in km/h. In logistic regression it is often seen that
accidents resulting in fatalities (0.3%), 216.1 resulting in severe injury the injury risk curves are not zero at zero impact speed. This reflects
(16.2%), and 1,119.9 resulting in minor injury (83.4%). Further ana- real life as the injuries may arise from the cyclist falling off the bicycle
lyses are all based on the weighted sample (Table 1). and impacting the ground rather than being struck by the vehicle

Table 1 Table 2
Weight factors for GIDAS-PCM data. Weight factors for injury risk curves using GIDAS data.
Fatal Severe Slight Σ Fatal Severe Slight Σ

Germany Accidents 382 14,124 65,222 79,728 Germany Accidents 382 14,124 65,222 79,728
(%) 0.48 17.72 81.81 100 (%) 0.48 17.72 81.81 100
GIDAS-PCM Accidents 11 481 1,470 1,962 GIDAS Accidents 54 1,891 6,655 8,600
(%) 0.56 24.52 74.92 100 (%) 0.63 21.99 77.38 100
Relative weight factors 1 0.85 1.28 Relative weight factors 1 1.06 1.39
Unweighted Sample Accidents 4 301 1,035 1,340 Unweighted Sample Accidents 16 545 2,101 2,662
(%) 0.3 22.46 77.24 100 (%) 0.6 20.47 78.93 100
Norm. weight factors 0.846 0.718 1.082 Norm. weight factors 0.757 0.803 1.053

4
H. Jeppsson and N. Lubbe Accident Analysis and Prevention 142 (2020) 105538

(Schramm, 2011).
The true positive rate (sensitivity) is the proportion of injured cy-
clists correctly classified while the false positive rate (1-specificity) is
the proportion of not-injured cyclist accidents that were incorrectly
classified. To summarize the true positive rate and false positive rate for
a data sample, a Receiver Operator Characteristic (ROC) curve was
created (Vittinghoff et al., 2005). The quality of the risk curves was
assessed by creating ROC curves using the repeated random sub-sam-
pling validation method. The training data consisted of 80% randomly
selected cases from the dataset and the testing data consisted of the
20% that was left. The process of splitting the data into training and
testing dataset and generating the ROC curve was repeated ten times.
For each ROC, an Area Under Curve (AUC) value was calculated. An
average AUC was then calculated for each severity. An average AUC of
1 indicates that the proportion of correctly classified cyclist injuries is
100%, while a value of 0.5 indicates that the proportion of correctly Fig. 2. Cyclist injury risk curves.
classified cyclist injuries is the same as the proportion of incorrectly
classified cyclist injuries and that the risk curve is no better than chance sparse and the curves overlap (Fig. 2). The injuries are not mutually
at predicting injury. The risk curves where further analysed calculating exclusive in our formulation, as they would have been for example in an
a Quality Index based on the relative size of the 95% confidence in- ordered probit model used by Kovaceva et al. (2020). In fact, a person
terval at three levels of risk: 5%, 25% and 50%. A value below 0.5 sustaining an AIS3+ injury is per definition also sustaining an AIS2+
indicates that the risk curves are good (have narrow confidence inter- injury. The 95% confidence intervals of the risk curves are depicted
vals), while a value above 1.5 indicates unacceptable risk curves (too together with the ROC curves in Appendix D.
wide confidence intervals) (Petitjean et al., 2012).
3.2. System effectiveness
2.7. Effectiveness
In terms of accidents avoided, increasing the baseline FoV from 75°
The effectiveness of the system in decreasing fatal, MAIS3+ and to 180° increases effectiveness by more than 20%. Similarly, adding the
MAIS2+ injuries was assessed using an established method based on VEB to the baseline 75° system (VEB 75°) increases avoidance effec-
Kullgren (2008). The braking system applied can reduce impact speed, tiveness by the same amount (Table 4). However, VEB 75° is in addition
and injury risk curves were used to calculate the non-exclusive risks of about 5%–18% more effective in reducing fatalities and injuries, and
being injured at the different severity levels with the new impact speed. thus adding VEB offers an advantage over increasing FoV in terms of
The effectiveness was specified according to Eq. (7), where N was the safety. The best performance is obtained by combining VEB with an
sum of the weighted injury probabilities of fatal, MAIS3+, and increased FoV: VEB 180° has an effectiveness of over 80% in preventing
MAIS2+ injuries in the sample and N’ was the sum of the weighted accidents and reducing injury at all levels of severity.
number of the estimated fatal, MAIS3+, and MAIS2+ injury prob-
abilities with the AEB or VEB system applied. 3.3. Characterization of remaining accidents
N'
E=1 VEB 75° was found to reduce vehicle impact speed markedly more
N (7)
than Baseline 180° with few remaining accidents having a vehicle im-
pact speed above 12 m/s (Fig. 3), a finding which explains VEB 75°’s
3. Results greater effectiveness in preventing injuries and fatalities. For cyclist
impact speed, the opposite can be seen with a greater reduction at
In this section we first present the injury risk curves, then quantify lower speeds and a smaller effect at higher speeds (Fig. 4). Neither of
system effectiveness, and finally analyse the remaining accidents, i.e. the AEB systems seems effective for cyclist speeds above 8 m/s. Fig. 5
those the systems were unable to prevent. illustrates the change in the distribution of the point of impact of the
cyclist. Initially, this point was fairly evenly distributed over the vehicle
3.1. Injury risk curves front, while in remaining accidents it is more concentrated around the
vehicle corners; both systems were found to be more effective in pre-
Cyclist injury risk curves for fatal, MAIS3+ and MAIS2+ injuries venting accidents in the centre of the vehicle front. Again, both systems
were calculated using Eq. (6). Regression coefficients are given in are very effective in accident scenarios where cyclists are crossing the
Table 3. Average AUC and Quality Index indicate reasonable to very path of a vehicle which is going straight or turning right. Additionally,
good prediction accuracy (Table 3). As expected, injury risk at each VEB 75° is highly effective for head-on accidents (Fig. 6). The dis-
severity increases with vehicle speed. Lower injury severities are more tribution of impact points and accidents scenarios was significantly
frequent at any given speed up to circa 90 km/h where data becomes different between systems (p-value in chi-square test < 0.001).

Table 3
Results from logistic regression, p-value, AUC and quality index.
Best estimate Lower limit Upper limit p-value AUC Quality Index 5% / 25% / 50%.

Fatal b0 −8.937 −10.560 −7.313 < 0.001 0.983 0.22 / 0.24 / 0.26
b1 0.113 0.083 0.143
MAIS3+ b0 −4.566 −4.949 −4.183 < 0.001 0.737 0.22 / 0.23 / 0.26
b1 0.064 0.052 0.076
MAIS2+ b0 −2.137 −2.312 −1.961 < 0.001 0.639 − / 0.19 / 0.25
b1 0.043 0.034 0.050

5
H. Jeppsson and N. Lubbe Accident Analysis and Prevention 142 (2020) 105538

Table 4 4. Discussion
System effectiveness for fatal, MAIS3+, MAIS2+ injury reduction (%) and
accident avoidance (%) in frontal car-to-cyclist accidents. The Baseline 75° system avoided 41% of all cyclist accidents, in line
Baseline 75° Baseline 180° VEB 75° VEB 180° with the earlier estimates of Zhao et al. (2019) and Rosén (2013) of
50% and 31%, respectively, when a FoV of 90° was used. VEB 75° has
Fatal 65 68 86 90 the same effectiveness as the Baseline 180° system at 59%, but with a
MAIS3+ 54 64 74 88
higher reduction in injuries of 67%–86% compared to 62%–68%.
MAIS2+ 47 62 67 85
Accidents avoidance 41 59 59 82 The VEB 180° system brings about the most significant reductions in
accidents, injuries and fatalities with an effectiveness exceeding 80%
for avoidance and 90% for fatality reduction. However, despite this
high performance, accidents still occur which indicates that in-crash
protection for cyclists is needed even when advanced AEB systems are
implemented. As the remaining accidents are more concentrated in the
corners of the vehicle front, protection at the A-pillar is of primary
importance. Fredriksson and Rosén (2014)’s conclusion that cyclist
head injuries are best prevented by a combination of AEB and de-
ployable hood lifters and external airbag seems justified even for the
most advanced AEB systems. Other measures such as helmets can also
contribute in their unique way to protecting the cyclist’s head from A-
pillar impacts. Notably, none of the fatally injured cyclists were
wearing a helmet in the data sample used to construct injury risk
curves; a lower injury risk at any given speed would be expected if all
cyclists wore helmets. To reduce cyclist head injuries, Ohlin et al.
(2017) suggest implementing a combination of AEB and VRU-friendly
vehicle front design alongside helmet usage, a suggestion which is
Fig. 3. Vehicle impact speed for all accidents and remaining accidents for supported by this study.
Baseline 180° and VEB 75°. Current Euro NCAP cyclist assessment covers accident scenarios
where the cyclist is crossing the vehicle path from the near-side and
where the cyclist and vehicle are travelling in the same direction.
However, given the distribution of accidents in GIDAS, this means that
only 35% of all vehicle-to-cyclist accidents are covered. Of the re-
maining accidents in Baseline 180°, more than half include a turning
vehicle, indicating the importance of adding such a scenario in future
evaluations of AEB systems. Since the AEB system in our study was
designed for the Euro NCAP scenarios, one would expect higher effec-
tiveness in these scenarios; but no such effect could be seen.
Half of the cyclists involved in accidents were travelling at higher
speeds than that tested for in the Euro NCAP scenario, which is 4.16 m/s
(15 km/h). However, modification of Euro NCAP scenarios planned for
2020 includes higher cyclist speeds, increasing the maximum speed from
15 km/h to 20 km/h, and lower car speeds, adding evaluations at
10 km/h (Euro NCAP, 2019b). If these changes are implemented, the
AEB settings then required to gain a full score would change: specifically,
Fig. 4. Cyclist impact speed for all accidents and remaining accidents for the trigger width would need to be increased to 3.1 m and the FoV to
Baseline 180° and VEB 75°. 130°: these numbers are calculated using Eqs. (1–5). A future baseline
AEB system such as this would clearly avoid more accidents and injuries.

Fig. 5. Distribution of impact point on vehicle front for all accidents and remaining accidents for Baseline 180° and VEB 75°.

6
H. Jeppsson and N. Lubbe Accident Analysis and Prevention 142 (2020) 105538

Fig. 6. Distribution of accident scenarios for all accidents and remaining accidents for Baseline 180° and VEB 75°.

Nonetheless, we estimate that VEB would still be around 20% more ef- When creating the risk curves, both the impact speed of the car (VK)
fective in accident avoidance and fatality prevention. and the relative speed between the car and the bicycle (VREL) were
investigated. Spitzhüttl and Liers (2016) suggested that VREL would be
advantageous over VK for cyclist risk curves, while Kovaceva et al.
4.1. Risk curves (2019) use VK for cyclist and pedestrian risk curves. Only small dif-
ferences between the two different impact speeds could be found for all
Injury risk curves were constructed to be dependent on impact three risk curves (Appendix E). The impact speed of the car (VK) was
speed only, while in reality many factors play a role in determining selected for our study for simplicity and comparability to Rosén (2013)
injury outcome, such as cyclist age and gender, vehicle type and design, and Kovaceva et al. (2019).
and use of protective equipment. These effects were not modelled as
they were not central to the research question and case numbers were
insufficient for more detailed analysis. 4.2. Strengths and limitations
Rosén (2013) created risk curves for fatal and severe injuries to
cyclists, and a comparison of Rosén’s risk curves and ours shows only The simulation model is a simplification of reality, using, for ex-
minor differences even though the sampling criteria differ: Rosén for ample, ideal sensors unaffected by adverse environmental conditions
example excludes cyclists under the age of 15 and does not restrict the such as difficult light or heavy rain. The data used for the study, GIDAS-
model year of the vehicle. In Rosén’s study, the fatal risk curve has an PCM, results in fewer sensor obstructions than in real traffic situations
intercept value, b0, of −8.8 and a slope, b1, of 0.098, compared to since other (moving) traffic participants are not recorded in the data-
−8.9 and 0.113 presented here. For the severe risk curve, Rosén’s in- base. Both these limitations might lead to the effectiveness of the AEB
tercept value is −4.7 and the slope value is 0.065, while our values are systems being overstated in the simulation compared to reality.
nearly identical at −4.6 and 0.064 respectively. In consequence, in However, the relative effects of comparing different AEB systems are
terms of injury reduction, the effectiveness of the systems presented less likely to be affected.
here would be found to be nearly identical whether using the presented In this study there was no restriction on the collision speed of the
injury risk curve or that of Rosén (2013); for fatal accidents, however, vehicle; all vehicle speeds were included. Euro NCAP tests car speeds
using the risk curves presented here would result in a marginally higher from 20 km/h to 60 km/h and accidents outside this speed range could
estimation of effectiveness. Another way of explaining the difference have been removed from the dataset to restrict the analysis to accidents
would be to calculate the exponential function of the slope value and addressed by current Euro NCAP testing. However, intentionally all
that would give the odds ratio (relative measure of effectiveness). Ro- accidents were considered as these are, in principle, addressable by
sén’s fatal risk curve has a lower odds ratio than ours, 1.10 compared to state-of-the-art AEB.
1.12. This means that for each 1 km/h increase in impact velocity, the On the other hand, only accidents where the cyclist impacts the
odds of sustaining an injury are 2% higher in our study compared to front of the vehicle were included in the study even though there are a
Rosén. large number of accidents in which the cyclist impacts the side of the
There were a few accidents at high speeds that did not lead to severe vehicle. These side impacts are more challenging to address by forward-
or fatal injuries. Investigating these accidents showed that only a part of looking sensors and brake intervention alone and were therefore out-
the cyclist’s front or rear wheel was hit by the vehicle, or that the cyclist side the scope of this study.
rolled up onto the car’s roof. These accidents may explain why the risk
of a MAIS2+ injury is lower than the risk of a fatal injury at speeds 5. Conclusion
higher than 90 km/h.
As can be seen in Fig. 2, the MAIS2+ risk curve is not zero at zero A simplified state-of-the art AEB system achieving the maximum
speeds. This is often seen in risk curves for slight injuries as these are score in Euro NCAP was evaluated through simulation in terms of how
recorded at very low speeds, and this issue can be addressed by simply successfully it avoids accidents with cyclists and reduces levels of cy-
defining the probability of injury as zero at zero speed (Schramm, clist fatality and injury. The baseline system was defined as having a
2011). FoV of 75° and activating not before 0.7 s TTC; this system avoided 41%

7
H. Jeppsson and N. Lubbe Accident Analysis and Prevention 142 (2020) 105538

of all accidents, 65% of fatalities and 47% of MAIS2+ injuries. Adding the most advanced AEB systems greatly reduce injuries yet cannot
more braking power by including the vacuum emergency brake (VEB) prevent them altogether.
resulted in 18% fewer accidents and a further reduction in both fatal- The remaining accidents were characterized by high cyclist speeds
ities and injuries of at least 20%. This was achieved with the same and impact to the vehicle corners, potentially increasing the relative
activation time and thus the same unnecessary activation rate as the frequency of head impacts to the stiff A-pillars. To best protect cyclists
baseline system. Increased crash avoidance performance and safety from injuries in impacts with passenger cars, advanced AEB systems
benefits can therefore be achieved without compromising unnecessary should therefore be combined with in-crash protection in the A-pillar
activation performance and nuisance activations. area.
Increasing the FoV to 180° in the baseline system increased accident
avoidance to the same level as when the VEB is added, indicating that
this is a relevant system improvement. However, fatality and injury CRediT authorship contribution statement
reduction remained 5%–18% lower compared to adding the VEB and
thus simply increasing FoV was less beneficial overall than increasing Hanna Jeppsson: Conceptualization, Methodology, Formal ana-
braking power. The combination of a wider, 180°, FoV combined with lysis, Writing - original draft, Writing - review & editing, Visualization.
the increased braking power of the VEB avoided 82% of accidents and Nils Lubbe: Conceptualization, Methodology, Writing - original draft,
90% of fatalities, a substantial increase over baseline, indicating that Writing - review & editing.

Appendix A. System parameters

Description of system parameters

FoV Sensor field-of-view


Reference: 75°. Variation: 180°
Rmin, Rmax Minimum and maximum range in which a VRU can be detected
Reference: Rmin = 0 m, Rmax = 60 m. No variations.
Frame rate Rate at which images from the sensor are gathered
Reference: 25 Hz. No variations.
tclassification Time from first possible detection of VRU until it has been detected and classified as pedestrian or bicyclist. Only after this time period can AEB be activated.
Reference: 3 frames = 120 ms. No variations
tlatency Processing time at each frame
Reference: 1 frame = 40 ms. No variations
amax Maximum brake acceleration provided by the AEB system. The acceleration in each particular accident was further constrained by the available tire‐to‐road friction.
Reference: g*μ.Variations:VEB: 1.8*g*μ
tramp Ramp‐up time from onset of brakes until max brake acceleration was provided
Reference: 150 ms. No variations.
TTCaeb Maximum predicted time to collision, for an unbraked car, when the brake decision was taken. Note that brake onset would occur slightly later due to tlatency.
Reference: 0.7 s. No variations.
w Trigger width. The maximum lateral distance from car path to VRU at which braking was activated.
Reference: 1.2 m. No variations.
adriver AEB is always activated regardless of how hard the driver brakes
vmax No cut‐off speed of car above which AEB was not activated
Darkness System active in darkness
Sensor position Center of gravity in longitudinal direction and middle of vehicle in lateral direction

Appendix B. brake time and TTC calculation

The energy an object has in motion is called kinetic energy and is dependent on the mass m and velocity v of the object. Work is needed to change
the object’s velocity. To reduce velocity to zero, all kinetic energy must be transformed, hence setting Work = kinetic energy can be used to calculate the
distance needed for the force to act, the brake distance.
d
F (s ) ds = Ekinetic (B1)
0

1 2
mad = mv
2 (B2)
With a being brake deceleration, a positive value. Solving for d, the brake distance:
v2
d=
2a (B3)
Using the basic kinematic equation for constant acceleration one can calculate the brake duration given the brake distance d:
1 2
d= at
2 (B4)

v 2
2d 2a 2v 2 v
t= = 2 = =
a a 2a2 a
(B5)
Lastly, TTC is defined as distance to the collision point divided by current speed for an unbraked vehicle, and TTC at can be expressed using brake
distance from Eq. B3:

8
H. Jeppsson and N. Lubbe Accident Analysis and Prevention 142 (2020) 105538

Fig B1. Example for a brake event.

v2
d v
TTC = = 2a =
v v 2a (B6)
In consequence, TTC at brake onset is equal to half the brake duration.
An example for a vehicle travelling at 10 m/s with an instantaneous deceleration of 10 m/s2 and an initial distance to the collision point of 10 m is
depicted Fig. B1. In order to come to a full stop (zero velocity) just before the collision point, the deceleration needs to start at 0.5 s after initial
conditions, that is 5 m prior to the collision point. At this time the TTC is 5 m divided by 10 m/s = 0.5 s. The vehicle slows down from 0.5 s to 1.5 s,
giving a brake time of 1.0 s. Hence, the TTC at brake onset is half the brake duration.

Appendix C. Brake distance calculation

The brake distance calculation which determines the distance it takes for the vehicle to come to a complete stop, dvehicle, consists of two parts, the
ramp-up phase and constant deceleration in braking phase (Fig. C1). Zero acceleration and constant speed at brake initiation are assumed. Fur-
thermore, acceleration increases linearly in the ramp-up time.
Ramp-up distance where t ∈ [0; t_{ramp}]
amax *t
a (t ) =
tramp (C1)
t amax *t 2
v (t ) = a (t ) dt + v0 = + v0
0 2*tramp (C2)

Fig. C1. The different phases of the brake event.

9
H. Jeppsson and N. Lubbe Accident Analysis and Prevention 142 (2020) 105538

t amax *t 3
d (t ) = v (t ) dt + v0 = + v0 *t
0 6*tramp (C3)
when t = tramp, then
amax * tramp2
dramp = + v0*tramp
6 (C4)
Constant deceleration braking distance, t > t_{ramp}
a = amax (C5)
Vehicle speed at the beginning of the constant deceleration phase
amax *tramp
vramp = + v0
2 (C6)
With Eq. C3, brake distance to come to a full stop under constant deceleration
amax * tramp
vramp2 ( 2
+ v0 ) 2 amax *tramp2 v0 2 v0*tramp
dconstant _ braking = = = +
2amax 2amax 8 2*amax 2 (C7)
Total braking distance (vehicle comes to a full stop)
d vehicle = dramp + dconstant _ braking = (C8)
=
amax * tramp2 amax * tramp2 v0 2 v0 * tramp
+ v0 * tramp + +
6 8 2 * amax 2

=
v0 2 v0 tramp amax tramp2
= +
2amax 2 24 (C9)

Appendix D. Injury risk curves and Receiver Operator Characteristic curves

Risk curves with 95 % confidence intervals for MAIS2+, MAIS3+ and fatal injuries are calculated using the formula
b 0 b1 v± 1.96(XTS2X)1/2)
conf.limits(v) = 1/(1 + e (D1)
where b0 and b1 are the resulting coefficients from the glm() function. X= (1, v) , and S2 also result from the logistic regression and show the variance-
covariance matrix.
Using the functions predict(), prediction() and performance() in R, a Receiver Operator Characteristic (ROC) curve can be created. For each ROC
the Area Under Curve (AUC) value is calculated. The black diagonal line shows where the true positive rate is equal to the false positive rate. A point
on this line means that proportion of correct classified injured cyclists is the same as the proportion of incorrect classified injured cyclists. A repeated
random sub-sampling validation method was used, and the ROC and AUC was calculated and repeated ten times. The training data consisted of 80%
of the dataset and the validation data consisted of 20% of the dataset (Figs. D1, D2, D3, D4, D5, D6 ).

Fig. D1. Risk of cyclist MAIS2+ injury with 95% confidence intervals and existence of MAIS2+ injury (gray dots) in a frontal accident with a modern passenger car.

10
H. Jeppsson and N. Lubbe Accident Analysis and Prevention 142 (2020) 105538

Fig. D2. ROC curves for MAIS2+ injury.

Fig. D3. Risk of cyclist MAIS3+ injury with 95% confidence intervals and existence of MAIS3+ injury (gray dots) in a frontal accident with a modern passenger car.

Fig. D4. ROC curves for MAIS3+ injury.

11
H. Jeppsson and N. Lubbe Accident Analysis and Prevention 142 (2020) 105538

Fig. D5. Risk of cyclist fatal injury with 95% confidence intervals and existence of fatal injury (gray dots) in a frontal accident with a modern passenger car.

Fig. D6. ROC curves for fatal injury.

Appendix E. Injury risk curves, comparison of VK and VREL

The risk curves for fatal, MAIS3+ and MAIS2+ injuries using relative impact speed between car and cyclist (VREL) are compared to the risk
curves using car impact speed (VK). ROC curves have also been calculated for VREL and show only small differences compared to using VK (Figs. E1,
E2, E3, E4, E5, E6 ).

Fig. E1. Comparison of risk using VK and VREL of cyclist MAIS2+ injuries in a frontal accident with a modern passenger car.

12
H. Jeppsson and N. Lubbe Accident Analysis and Prevention 142 (2020) 105538

Fig. E2. ROC curves for VREL and MAIS2+ injury. Dataset divided into training data (80%) and validation data (20 %) sets.

Fig. E3. Comparison of risk using VK and VREL of cyclist MAIS3+ injuries in a frontal accident with a modern passenger car.

Fig. E4. ROC curves for VREL and MAIS3+ injury. Dataset divided into training data (80%) and validation data (20 %) sets.

13
H. Jeppsson and N. Lubbe Accident Analysis and Prevention 142 (2020) 105538

Fig. E5. Comparison of risk using VK and VREL of cyclist fatal injuries in a frontal accident with a modern passenger car.

Fig. E6. ROC curves for VREL and fatal injury. Dataset divided into training data (80%) and validation data (20%) sets.

Appendix G. Supplementary data

Supplementary material related to this article can be found, in the online version, at doi:https://doi.org/10.1016/j.aap.2020.105538.

References Document]. February. URL https://cdn.euroncap.com/media/43371/euro-ncap-


aeb-vru-test-protocol-v301.pdf (accessed 10.14.19). .
Euro NCAP, 2017. European new car assessment programme, assement protocol for pe-
Ames, E., Martin, P., 2015. Pop-up hood pedestrian protection. In: 24th International destrian protection [WWW Document]. URL https://cdn.euroncap.com/media/
Enhanced Safety of Vehicles (ESV) Conference. Gothenburg, Sweden. 32282/euro-ncap-assessment-protocol-pp-v902.pdf (accessed 10.14.19). .
Edwards, M., Nathanson, A., Wisch, M., 2014. Estimate of potential benefit for Europe of European Commission, 2016. Vehicle Safety 2016 [WWW Document]. URL https://ec.
fitting Autonomous Emergency Braking (AEB) systems for pedestrian protection to europa.eu/transport/road_safety/sites/roadsafety/files/ersosynthesis2016-
passenger cars. Traffic Inj. Prev. 15 (Suppl 1), S173–82. https://doi.org/10.1080/ vehiclesafety15_en.pdf (accessed 10.14.19). .
15389588.2014.931579. Fredriksson, R., Håland, Y., Jikuang, Y., 2001. Evaluation Of A New Pedestrian Head
Elvik, R., Mysen, A.B., 1999. Incomplete accident reporting, meta-analysis of studies Injury Protection System With A Sensor In The Bumper And Lifting Of The Bonnet’s
made in 13 countries. Transp. Res. Rec. 1665 (1), 133–140. https://doi.org/10.3141/ Rear Part. In: 17th International Technical Conference on the Enhaced Safety of
1665-18. Vehicles (ESV). Amsterdam, The Netherlands. pp. 1–12.
Erbsmehl, C., 2009. Simulation of real crashes as a method for estimating the potential Fredriksson, R., Ranjbar, A., Rosen, E., 2015. Integrated Bicyclist Protection Systems -
benefits of advanced safety technologies. In: 21st International Technical Conference Potential of Head Injury Reduction Combining Passive and Active Protection
on the Enhaced Safety of Vehicles (ESV). Stuttgart, Germany. Systems. In: 24th International Enhanced Safety of Vehicles (ESV) Conference.
Esmaeilikia, M., Radun, I., Grzebieta, R., Olivier, J., 2019. Bicycle helmets and risky Gothenburg, Sweden. pp. 1–12.
behaviour: A systematic review. Transp. Res. Part F Traffic Psychol. Behav. 60 Fredriksson, R., Rosén, E., 2014. Head Injury Reduction Potential of Integrated Pedestrian
(January), 299–310. https://doi.org/10.1016/j.trf.2018.10.026. Protection Systems Based on Accident and Experimental Data – Benefit of Combining
Euro NCAP, 2019a. Euro NCAP Test Protocol - AEB VRU systems [WWW Document]. Passive and Active Systems. In: Proceedings of IRCOBI Conference. Berlin, Germany,
URL https://cdn.euroncap.com/media/43381/euro-ncap-aeb-vru-test-protocol- p. IRC-14-69.
v204.pdf (accessed 10.14.19). . Haddon, W., 1970. On the Escape of Tigers: An Ecologic Note. Am. J. Public Health 60
Euro NCAP, 2019b. European New Car Assessment Programme (Euro NCAP) [WWW

14
H. Jeppsson and N. Lubbe Accident Analysis and Prevention 142 (2020) 105538

(12), 6–7. 10.18.19). .


Held, F., 2016. Investigation of under-reporting and the consistency of injury severity Petitjean, A., Trosseille, X., Praxl, N., Hynd, D., Irwin, A., 2012. Injury Risk Curves for the
classifications in Swedish police crash data compared to hospital injury data based on WorldSID 50 th Male Dummy. Stapp Car Crash J. 56 (October), 323–347. https://doi.
the Swedish Traffic Accident Data Acquisition (STRADA). org/10.4271/2012-22-0008.
Hu, J., Klinich, K.D., 2012. TOWARD DESIGNING PEDESTRIAN-FRIENDLY VEHICLES. Ranjbar, A., 2014. Active Safety for Car-to-Bicyclist Accidents.
Isaksson-Hellman, I., Werneke, J., 2017. Detailed description of bicycle and passenger car Rosén, E., 2013. . Autonomous Emergency Braking for Vulnerable Road Users. In:
collisions based on insurance claims. Saf. Sci. 92, 330–337. Proceedings of IRCOBI Conference. Gothenburg, Sweden. pp. 618–627.
ITF, 2019. Road Safety in European Cities: Performance Indicators and Governance Rosén, E., Källhammer, J.E., Eriksson, D., Nentwich, M., Fredriksson, R., Smith, K., 2010.
Solutions. International Transport Forum Policy Papers. OECD Publishing, Paris. Pedestrian injury mitigation by autonomous braking. Accid. Anal. Prev. 42 (6),
James, S.L., Abate, D., Abate, K.H., Abay, S.M., Abbafati, C., et al., 2017. Global, regional, 1949–1957. https://doi.org/10.1016/j.aap.2010.05.018.
and national incidence, prevalence, and years lived with disability for 354 Diseases Roth, G.A., Abate, D., Abate, K.H., Abay, S.M., Abbafati, C., et al., 2018. Global, regional,
and Injuries for 195 countries and territories, 1990-2017: A systematic analysis for and national age-sex-specific mortality for 282 causes of death in 195 countries and
the Global Burden of Disease Study 2017. Lancet 392, 1789–1858. https://doi.org/ territories, 1980–2017: a systematic analysis for the Global Burden of Disease Study
10.1016/S0140-6736(18)32279-7. 2017. Lancet 392 (392), 1736–1788. https://doi.org/10.1016/S0140-6736(18)
Jeppsson, H., Östling, M., Lubbe, N., 2018. Real life safety benefits of increasing brake 32203-7.
deceleration in car-to-pedestrian accidents: Simulation of Vacuum Emergency Sander, U., Lubbe, N., 2018. Market penetration of intersection AEB: Characterizing
Braking. Accid. Anal. Prev. 111 (December 2017), 311–320. https://doi.org/10. avoided and residual straight crossing path accidents. Accid. Anal. Prev. 115,
1016/j.aap.2017.12.001. 178–188. https://doi.org/10.1016/j.aap.2018.03.025.
Katsuhara, T., Miyazaki, H., Kitagawa, Y., Yasuki, T., 2014. Impact Kinematics of Cyclist Schramm, S., 2011. Methode zur Berechnung der Feldeffektivität integraler
and Head Injury Mechanism in Car-to-Bicycle Collision. In: Proceedings of IRCOBI Fußgängerschutzsysteme.
Conference. Berlin, Germany. pp. 670–684. Silla, A., Leden, L., Rämä, P., Scholliers, J., Van Noort, M., Bell, D., 2017. Can cyclist
Kovaceva, J., Bálint, A., Schindler, R., Schneider, A., 2020. Safety benefit assessment of safety be improved with intelligent transport systems? Accid. Anal. Prev. 105,
autonomous emergency braking and steering systems for the protection of cyclists 134–145. https://doi.org/10.1016/j.aap.2016.05.003.
and pedestrians based on a combination of computer simulation and real-world test Spitzhüttl, F., Liers, H., 2016. Methodik zur Erstellung von Verletzungsrisikofunktionen
results. Accid. Anal. Prev. 136 (October 2019), 105352. https://doi.org/10.1016/j. aus Realunfalldaten [WWW Document]. URL https://www.vda.de/de/services/
aap.2019.105352. Publikationen/Publikation.∼1492∼.html (accessed 10.18.19). .
Kovaceva, J., Bálint, A., Schindler, R., Schneider, A., Johann, S., et al., 2019. PROSPECT - Statistisches Bundesamt (Destatis), 2018a. Kraftrad-und Fahrradunfälle im
Proactive Safety for Pedestrian and Cyclists. Straßenverkehr 2017 [WWW Document]. URL http://lauf-an-der-pegnitz.info/files/
Kullgren, A., 2008. Dose-response models and EDR data for assessment of injury risk and unfaelle-zweirad-5462408177004.pdf (accessed 10.18.19). .
effectiveness of safety systems. In: Proceedings of IRCOBI Conference. Bern, Statistisches Bundesamt (Destatis), 2018b. Verkehr Verkehrsunfälle 2017 [WWW
Switzerland. pp. 3–14. Document]. URL https://www.destatis.de/DE/Themen/Gesellschaft-Umwelt/
Miljö-och energidepartementet, 2018. Strategi för Levande städer – politik för en hållbar Verkehrsunfaelle/Publikationen/Downloads-Verkehrsunfaelle/verkehrsunfaelle-
stadsutveckling [WWW Document]. URL https://www.regeringen.se/4971fa/ jahr-2080700177004.pdf?__blob=publicationFile&v=4 (accessed 10.18.19). .
contentassets/b5640fd317d04929990610e1a20a5383/171823000webb.pdf (ac- Stigson, H., Fahlstedt, M., 2016. Preventing Shoulder Injuries in Bicycle Crashes. In:
cessed 10.18.19).. . International Cycling Safety Conference. Bologna, Italy.
Näringsdepartementet, 2009. Mål för framtidens resor och transporter, Prop. 2008/09:93 Trafikanalys, 2018. Vägtrafikskador [WWW Document]. URL. https://www.trafa.se/
[WWW Document]. URL https://www.regeringen.se/49bbc2/contentassets/ vagtrafik/vagtrafikskador/.
80dd7d80fc64401ca08b176a475393c5/mal-for-framtidens-resor-och-transporter- van Ratingen, M., Williams, A., Lie, A., Seeck, A., Castaing, P., et al., 2016. The european
prop.-20080993 (accessed 10.18.19). . new car assessment programme: A historical review. Chin. J. Traumatol. – Eng. Ed.
OECD/ITF, 2018. Road Safety Annual Report 2018 [WWW Document]. URL https:// 19 (2), 63–69. https://doi.org/10.1016/j.cjtee.2015.11.016.
www.itf-oecd.org/sites/default/files/netherlands-road-safety.pdf (accessed Vittinghoff, E., Glidden, D.V., Shiboski, S.C., McCulloch, C.E., 2005. Regression Methods
10.18.19). . in Biostatistics. Springer Science+Business Media, Inc.
Ohlin, M., Strandroth, J., Tingvall, C., 2017. The combined effect of vehicle frontal de- VUFO (Institute for Traffic Accident Research at Dresden University of Technology),
sign, speed reduction, autonomous emergency braking and helmet use in reducing 2019. Pre-Crash-Matrix (PCM) Format Specification v5.0. Dresden. Germany. .
real life bicycle injuries. Saf. Sci. 92, 338–344. https://doi.org/10.1016/j.ssci.2016. Wegman, F., Zhang, F., Dijkstra, A., 2012. How to make more cycling good for road
05.007. safety? Accid. Anal. Prev. 44 (1), 19–29. https://doi.org/10.1016/j.aap.2010.11.010.
Olivier, J., Creighton, P., 2017. Bicycle injuries and helmet use: A systematic review and Wood, J.M., Lacherez, P.F., Marszalek, R.P., King, M.J., 2009. Drivers’ and cyclists’ ex-
meta-analysis. Int. J. Epidemiol. 46 (1), 278–292. https://doi.org/10.1093/ije/ periences of sharing the road: Incidents, attitudes and perceptions of visibility. Accid.
dyw153. Anal. Prev. 41 (4), 772–776. https://doi.org/10.1016/j.aap.2009.03.014.
ONISR, 2017. La sécurité routière en France Bilan de l’accidentalité de l’année 2017 Zhao, Y., Ito, D., Mizuno, K., 2019. AEB effectiveness evaluation based on car-to-cyclist
[WWW Document]. URL https://cache.media.eduscol.education.fr/file/Securite_ accident reconstructions using video of drive recorder. Traffic Inj. Prev. 20 (1),
Routiere/95/3/bilan_ONISR_2017_accidentalite_routiere_1096953.pdf (accessed 100–106. https://doi.org/10.1080/15389588.2018.1533247.

15

You might also like