반도체디스플레이기술학회지 제19권 제3호(2020년 9월)

Journal of the Semiconductor & Display Technology, Vol. 19, No. 3. September 2020.

Improved Crash Detection Algorithm for Vehicle Crash Detection

†
Byoungman An* and YoungSeop Kim*
*† Electronic and Electrical Engineering, Dankook University, Korea Dankook

ABSTRACT
A majority of car crash is affected by careless driving that causes extensive economic and social costs, as well as
injuries and fatalities. Thus, the research of precise crash detection systems is very significant issues in automotive
safety. A lot of crash detection algorithms have been developed, but the coverage of these algorithms has been limited
to few scenarios. Road scenes and situations need to be considered in order to expand the scope of a collision
detection system to include a variety of collision modes. The proposed algorithm effectively handles the x, y, and z
axes of the sensor, while considering time and suggests a method suitable for various real worlds. To reduce nuisance
and false crash detection events, the algorithm discriminated between driving mode and parking mode. The
performance of the suggested algorithm was evaluated under various scenarios, and it successfully discriminated
between driving and parking modes, and it adjusted crash detection events depending on the real scenario. The
proposed algorithm is expected to efficiently manage the space and lifespan of the storage device by allowing the
vehicle's black box system to store only necessary crash event’s videos.
Key Words : Crash Detection, Automotive, Crash Event, Black Box, Video Recording

1. Introduction 1
system that helps reduce fatalities from traffic accidents. A
variety of new sensors have been used in the past few years
Vehicle crashes cause injuries and deaths to road users as for tasks such as recognition, navigation and manipulation
well as enormous economic and social costs, and more than of objects. [2]
50% of vehicle crashes are caused by careless driving. [1]. A plenty of technology has been aimed to research the
In order to ensure driving safety, in recent years, all crash detection in vehicle. Collision probability data
countries have researched car crash avoidance technology. generated from Monte Carlo simulation taking driver
According to statistics, if you can give the driver an behavior and vehicle dynamics into account, tracking
additional 0.5 second reaction time in a dangerous situation, algorithm using interactive multi-model particle filter, and
you can reduce collision by 45%, so modern cars are threat assessment algorithm to estimate collision probability
equipped with all kinds of measurement and alarm systems [2]. In another method, two models are considered: a model
to keep driving safety. Vehicle crash detection can be in which the follower maintains a safe distance and a model
detected in a variety of ways, but g-sensors (accelerometer) in which the follower maintains a safe time. Analyze
are usually used. This method is commonly used because it distance delays and time delays caused by major vehicles'
is cheap and possible to easily determine the amount of impact on followers. [3] There is also a way to develop new
change in the values of x, y, z. Today, being able to support challenging benchmarks for stereo, optical flow, visual
humans in their daily work by developing autonomous odometer/SLAM and 3D object detection tasks using
systems is one of the biggest challenges of modern autonomous driving platforms. [4]
computer science. One example is an autonomous driving The motion sensor method uses a complex motion
†
processor to provide very accurate data and, if used near the
E-mail: [email protected]
engine, can also filter out vibrations. They used automotive

93
94 Byoungman An · YoungSeop Kim

+y sensor [13], detecting the precise location of a moving car

+z by using a design system of the gyro sensor [14] and
investigating experimental security analysis of a modern
automobile. [15]
-x +x In this paper, A novel algorithm for vehicle crash
detection both driving and parking mode is presented. The
purposed system concept includes a crash detection and
safety distance calculation. The system detects the distance
-z between the vehicle and the vehicle front (object) and uses
-y vehicle CAN signal information from other devices.
Moreover, by considering the situation over time, we
Fig. 1. G-sensor (accelerometer) axes.
devised a shock event algorithm that is more suitable for the
real world. This paper is structured as follows: Section 2
sensors to reduce vibration and made readings very accurate. describes the background with the major concept. Section 3
Algorithms that use data fusion between acceleration, presents our preliminary work as well as purposed idea.
deceleration, and tilt angles have a great success rate. No Section 4 presents the experimental results from purposed
false positives or failed crashes were recorded in the test designs. Finally, Section 5 concludes the paper before
results. Collision detection is a very important feature for proposing the future of study.
motorcycle occupants and is used by the E-call system,
which can reduce the time between an accident and
2. Background
emergency service arrival by 50%. [5]
There is also a method that uses the image to recognize There are various methods for determining the crash
the license plate of the vehicle using Automatic license plate event, using images or using autonomous driving techniques,
recognition (ALPR). [6] The distance is calculated from but in this paper, we propose a method using g-sensor,
geometric derivation using additional descriptive data such which is a basic method.
as the distance between the cameras and other specific
+y,
angles, such as the angle of view of the camera. This method
1.000 +z, 1.000
achieves very high accuracy if the calculated distance
between the vehicle and the camera is relatively accurate.
The importance of measuring vehicle-to-vehicle distances -x, -1.000 +x, 1.000
lies in performing some important tasks in automotive
vehicle systems, such as vehicle speed calculations and
decision support (e.g. vehicle bypass, route change and
speed control). In general, the work of the literature is
-z, -
divided into two types of distance measuring systems based 1.000 -y, -
on image processing: mono vision systems and stereo vision 1.000
systems. [7] Fig. 2. Vehicle layout of g-sensor, the initial value is
In addition, there are algorithms using the ultrasonic specified as 1.000 for convenience and
sensor [8], method using vehicle-mounted device-based understanding of calculation.
collision risk identification and warning system [9], pre-
collision hot spot detection method using in-vehicle 2.1 G-sensor
odometer recorder [10], magnetic resistance and sonar The g-sensor is commonly referred to as an accelerometer.
sensors to detect impending collisions in cars [11], a They are used in various devices such as smartphones,
distance obstacle detection and safety distance calculation vehicles and of course black boxes. The black box's g-sensor
[12], adopting position control of an object using vision monitors the appropriate acceleration called G-Force. The 3-

반도체디스플레이기술학회지 제19권 제3호, 2020

 Improved Crash Detection Algorithm for Vehicle Crash Detection 95

axis accelerometer contains 3 accelerometers, one for each between the values of the gyroscope and accelerometer.
axis, which can measure the acceleration in the ±x, ±y, ±z,
axes as shown in Fig. 1. The accelerometer output is highly angle = 0.90 x (angle + gyroData x Δt)
dependent on the selected sensitivity expressed as G-force. + 0.10 x accData (1)
For example, an accelerometer with 2G sensitivity can
output an acceleration of up to 2G. The value is read from 2 The weight chosen for the data coming from both sensors
bytes register and the precision when using high sensitivity depends on the target application. Increasing the weight of
are sufficient for the consumer to use. Decreasing the the accelerometer data improves the responsiveness of the
accelerometer's sensitivity also decreases precision because system, but increases its sensitivity to vibration, which
the length of the register where the value is stored is the makes the system unstable. If the gyro weighs more than
same. There are 3 registers where acceleration is stored, X 0.90, it makes a slow but very stable system. [5]
acceleration register, Y acceleration register and Z
acceleration register. The data collected on each axis is 2.3. Car Crash Detection System
averaged and the values obtained are used to offset the As shown in Fig. 2, the arrangement of g-sensors in a
output each time the sensor is read. Most accelerometers vehicle may vary depending on the vehicle's space and
have offset registers, and writing the values obtained after design. In general, the black box is placed in the center of
calibration into these registers will offset the output. the vehicle because it is considered to be placed in a safe
For instance, the accelerometer's digital output has a full location. In general, g-sensor uses the first value read after
programmable range of ±2G, ±4G, ±8G and ±16GYou can system booting as the default value and measures the change
select the appropriate sensitivity according to the application. of the corresponding value. In this paper, the initial value is
In this paper, STMicroelectronics g-sensor was used and 2G specified as 1.000 for convenience and understanding of
was set up for measuring the accelerations of a vehicle. calculation. There are countless crash detection algorithms.
In this paper, the method of detecting impact through
2.2 Raw Data absolute values is as follows.
The raw values of the accelerometer are read by the
microcontroller and are obtained by using complementary |x2| – |x1| > x axis threshold
filters to perform data fusion. In this case, the accelerometer |y2| – |y1| > y axis threshold (2)
is used to correct the drift of the gyroscope. Complementary |z2| – |z1| > z axis threshold
filter is an equation that creates a weighted arithmetic mean

G-sensor
(Accelerometer)
(x, y, z) raw data

Data Collection
(Sampling rate 10ms)

Data Buffer #2
(x2, y2, z2)
Purposed Purposed
Crash
Sensor Crash
Event
Calibration Detection
Data Buffer #1
(x1, y1, z1)

Fig. 3. Architecture of the proposed system.

Journal of KSDT Vol. 19, No. 3, 2020

96 Byoungman An · YoungSeop Kim

The G-sensor value can reflect the period by sampling x1 = (1st avg. + 2nd avb.) / 2 (5)
that is the changeable value. In this paper, the sampling
period is 10ms for testing. x1, y1, z1 are the previous values, The processing of the value calculated in the above
and x2, y2, z2 are the changed values after 10ms. When the equation will be dealt with in detail in the algorithm
previous value is subtracted from the changed value as in proposed in Section 3.2.
Equation 2, if the corresponding value is higher than each
predetermined threshold value, it can be determined as the 3.2 Purposed Crash Detection
crash event. We will deal with it in more detail with the The flow chart of the proposed crash detection algorithm
algorithm suggested in the chapter 3. is shown in Fig. 4. The proposed algorithm calculates the

3. A Novel Algorithm of Vehicle Crash Start

Detection
System initialization
In this section, we present a novel algorithm for vehicle
crash detection algorithm both driving mode and parking
mode. Fig. 3 shows the architecture of the proposed system Set sampling rate (10ms)
of vehicle crash event determining.
Data collection G-sensor
3.1 Purposed Sensor Calibration
Accurate sensor calibration is key to obtain reliable
Sensor Calibration
ground information. The calibration pipeline proceeds as
follows: we calibrate the g-sensor intrinsically and
extrinsically and rectify the input data. Existing techniques Crash
Detection
for this work were necessary because they were not accurate
enough to calculate ground-based estimates. [4] Vector sum calculation
In this paper, we suggest a method to increase the
accuracy of data. The sample rate plays a significant role in
Compare between vector sum and
collecting accurate data. If the sample rate is low, the
the threshold
amount of collected data is increased as well as it causes
computational complexity. On the other hand, On the other
No
hand, if the sample rate is high, it is possible to miss a if, vSUM >
moment of crash event. The proposed method is presented vTHRD?
in order to solve this problem as following. First, calculate
Yes
the average sampling value of continuous sampling data as
No
shown in Equation (3). Second, calculate the average of
if, t > tTHRD?
each odd and even sampled data as shown in Equation (4).
Finally, by calculating the average of the two values as
Yes
Equation (5), more reliable data were obtained than using
each method (3, 4). The example below is a formula for
calculating x1. In this way, the values of y1, z1 and x2, y2, z1
Crash Event
can be obtained.

1st avg. : (X1 + X2 + X3 + X4 + X5 + X6) / 6 (3) End

2nd avg. : (X1 + X3 + X5) / 3 + (X2 + X4 + X6) / 3 (4)

Fig. 4 Purposed flow chart of crash detection

반도체디스플레이기술학회지 제19권 제3호, 2020

 Improved Crash Detection Algorithm for Vehicle Crash Detection 97

vector sum of x, y, z and then calculates the difference Table 1. Crash detection comparison
between the previous value and the current value. In Scenario Direction C#1 C#2 Proposed
addition, we propose a more accurate crash detection F DET DET DET
method by considering the time at which the accumulated FR DET DET DET
data were acquired, not only the instantaneous time when a
R1 DET - DET
crash is detected.
R2 DET - DET
First, the equation for obtaining the vector sum and the
difference between the previous and current values is as RR - DET DET
follows. R DET DET DET
Driving
LR - DET DET
vSum = √((x2 - x1)2 - (y2 - y1)2 - (z2 - z1)2) (6) L2 - DET DET
vSum > vTHRD  A possibility of crash
L1 DET - DET
FL - DET DET
where, vSum = the vector sum
x2, y2, z2 = the current calibrated values T DET DET -
x1, y1, z1 = the previous calibrated values B DET DET DET
F DET DET DET
if, vSum > vTHRD, it could be considered that there is a FR - - DET
probability of crash in vehicle. vTHRD is a threshold values R1 DET DET DET
of vector sum. The vTHRD may vary depending on the
R2 - - DET
vehicle environment. In this paper, the result of reviewing
RR - DET DET
the vector sum of generally generated shocks is to be based
on a value of about 2. The reason why it was selected as 2 in R - - DET
Parking
Chapter 4 was mentioned in detail. In a general situation, the RL DET DET DET
impact was determined based on the corresponding value, L2 - DET DET
and it is expected that it can be used by tuning according to L1 DET - DET
the environment of the vehicle based on the value.
FL - DET DET
Time is always important. In many systems, time is
T DET DET -
considered, but most do not. Considering time, the system
can be complicated, so it can be a poison rather than a B - - -
benefit. In this paper, a simple relationship between time and DET : Crash Detected
impact was considered so as not to affect the design of the - : Not detected
system as much as possible. It was a simple method, but C#1 : Commercial black-box device #1
what was considered and not considered showed a big C#2 : Commercial black-box device #2
difference in the results.

where,
if ((vSum > vTHRD) && (t > tTHRD))
t = a crash detected time of vSUM > vTHRD
{
tTHRD = sampling rate * calibration count * n
Crash Detection  Crash Event (7)
tTHRD can be determined in various ways by n, and the
}
calibration count is the amount of data used for the average
used in equation (3,4,5). For example, if the sampling rate is
10ms, the calibration count is 6, and n is 10, the value of
tTHRD can be 600ms. As confirmed through the experiment,
it was confirmed that there was a considerable influence on

Journal of KSDT Vol. 19, No. 3, 2020

98 Byoungman An · YoungSeop Kim

R 4.2 Crash Detection Comparison

RL RR The result of comparing the proposed algorithm and two
commercial products are described in the Table 1. The
experiment was carried out in the same environment for a
fair test, and each experiment was performed 10 times for an
L2 T R2
accurate experiment, and the result by the average is
indicated in the Table 1. The reason vTHRD mentioned in
B Section 3.2 was set to 2 is that when the initial value
L1 R1 calibrated in most of the tests was 1.0. When the car is hit by
a tennis ball that the value was around 2.0. Experiments are
possible in a variety of ways, but in this paper, for safety, the
vehicle is driving less than 10km during the driving test. In
the test, a repeated test was conducted in which a tennis ball
FL FR
was lightly thrown and hit in the driving and parking state in
F each of the F (Front), FR(Front Right), R1 (Right Front
Fig. 5. Direction of vehicle crash detection Door), R2 (Right Rear Door), RR (Rear Right), R (Rear),
during driving and parking mode RL (Rear Left), L1 (Left Front Door), L2 (Left Rear Door),
FL (Front Left), T (Top), and B (Bottom). During driving,
the bottom was tested over the bump, and during parking
the values, and when the proposed method was applied, the was omitted. Looking at the experimental results in Table 1,
impact could be judged more accurately than before. the proposed algorithm was able to detect impact in most
If both of conditions were satisfied, it was determined as directions of the overall average except for the top case. On
crash detection. The determined crash information occurred the other hand, in the case of commercial products, when an
as an event and made it possible to process in the system. In impact is applied from a specific direction, there are many
this paper, when a crash event occurs, the system operates cases where the impact is relatively not recognized. In the
video recording, a system log file, an alarm sound to check proposed algorithm, the top crash event could be recognized
whether it occurs, or not. by lowering the vTHRD value in the experimental vehicle.
The next chapter 4 describes the experimental results of However, unnecessary shaking could be recognized as a
the method proposed in this algorithm. crash in other points. Since the experiment considered the
overall performance, tuning for the top was not possible, but
4. Experimental Result if vTHRD and tTHRD are obtained through many
experiments, it is expected that the impact on the entire area
4.1 Performance Evaluation
can be sufficiently detected. Overall, there was an
The performance of the proposed crash detection
improvement in the proposed method, and it is necessary to
algorithm was evaluated for two scenarios via offline
experiment in various vehicle environments.
simulations, including crash of various direction (see Fig. 5).
To show the benefit of the proposed algorithm, driving,
parking scenarios were considered. The driving scenario 5. Conclusion
was considered to compare the crash detection between the This paper has proposed a novel crash detection algorithm
proposed algorithm and a general black-box available in a applicable to general driving and parking mode scenarios
market. The test environment was compared after installing that consider a various direction of car and that extracts as
the product proposed to the test vehicle and the product used, much information as possible from the calibrated data. The
and the crash was applied in the actual road and parking algorithm successfully detected in various directions of
environment to determine whether the crash was judged. crash. It is expected that the proposed crash detection
algorithm is reliable although further evaluations are

반도체디스플레이기술학회지 제19권 제3호, 2020

 Improved Crash Detection Algorithm for Vehicle Crash Detection 99

required. The proposed algorithm can be used as an “Automatic License Plate Recognition (ALPR): A State-
integrated collision detection algorithm by integrating of-the-Art Review,” in IEEE Transactions on Circuits
and Systems for Video Technology, vol. 23, no. 2, pp.
tracking information from multiple sources for collision
311-325, Feb. 2013.
warning, avoidance and mitigation. Throwing new light on 7. A. Zaarane, I. Slimani, A. Hamdoun and I. Atouf,
existing methods, we hope that the proposed algorithm will “Vehicle to vehicle distance measurement for self-
complement others and help to reduce overfitting to datasets driving systems,” 2019 6th International Conference on
with little test examples or training as well as contribute to Control, Decision and Information Technologies
(CoDIT), Paris, France, 2019, pp. 1587-1591.
the development of algorithms that work well in the crash
8. S. Haria, S. Anchaliya, V. Gala and T. Maru, “Car Crash
detection system of vehicle. Prevention and Detection System Using Sensors and
Smart Poles,” 2018 Second International Conference on
Acknowledgements Intelligent Computing and Control Systems (ICICCS),
Madurai, India, 2018, pp. 800-804.
This project would not have been possible without the 9. Q. Zeng, C. Wu, L. Peng and H. Li, “Novel vehicle
kind support of many individuals. I would like to extend our crash risk detection based on vehicular sensory system,”
sincere thanks to all of them. In addition, I would like to 2015 IEEE Advanced Information Technology, Electronic
and Automation Control Conference (IAEAC),
express our special gratitude and thanks to Management of
Chongqing, 2015, pp. 622-626.
Prof. Youngseop Kim for facilitating me by various 10. J. Lee and K. Jang, “Proactive Detection of Crash
resources and sufficient time whenever needed. Hotspots Using In-Vehicle Driving Recorder,” 2016 3rd
Asia-Pacific World Congress on Computer Science and
Engineering (APWC on CSE), Nadi, 2016, pp. 193-198.
References 11. S. Taghvaeeyan and R. Rajamani, “Two-Dimensional
1. R. R. Knipling et al., “Assessment of IVHS Sensor System for Automotive Crash Prediction,” in
countermeasures for collision avoidance systems,” IEEE Transactions on Intelligent Transportation Systems,
National Highway Traffic Safety Administration, vol. 15, no. 1, pp. 178-190, Feb. 2014.
Washington, DC, USA, Tech. Rep. DOT HS 807 995, 12. Y. Chen and C. Wang, “Vehicle Safety Distance
May 1993. Warning System: A Novel Algorithm for Vehicle Safety
2. T. Kim and H. Jeong, “A Novel Algorithm for Crash Distance Calculating Between Moving Cars,” 2007
Detection Under General Road Scenes Using Crash IEEE 65th Vehicular Technology Conference -
Probabilities and an Interactive Multiple Model Particle VTC2007-Spring, Dublin, 2007, pp. 2570-2574.
Filter,” in IEEE Transactions on Intelligent Transportation 13. E. Ha, G. Choi, “Position Control of an Object Using
Systems, vol. 15, no. 6, pp. 2480-2490, Dec. 2014. Vision Sensor,” KSDET, vol.10, no.2, 25, 2011, pp. 49-
3. V. Naumov, “Analysis of Time and Distance Delays in 56.
Car Following Models,” 2010 International Conference 14. M. Kim, J. Kim, O, Yang, “Design of the Position
on Intelligent Systems, Modelling and Simulation, Control System for Parabolic Antenna using Gyro
Liverpool, 2010, pp. 296-299. Sensor,” KSDET, vol. 12, no.2, 43, 2013, pp.85-91.
4. Geiger, P. Lenz and R. Urtasun, “Are we ready for 15. K. Koscher et al., “Experimental Security Analysis of a
autonomous driving? The KITTI vision benchmark Modern Automobile,” 2010 IEEE Symposium on
suite,” 2012 IEEE Conference on Computer Vision and Security and Privacy, Berkeley/Oakland, CA, 2010, pp.
Pattern Recognition, Providence, RI, 2012, pp. 3354- 447-462, doi: 10.1109/SP.2010.34.
3361.
5. A. Cismas, M. Ioana, C. Vlad and G. Casu, “Crash
Detection Using IMU Sensors,” 2017 21st International
Conference on Control Systems and Computer Science
(CSCS), Bucharest, 2017, pp. 672-676. 접수일: 2020년 9월 6일, 심사일: 2020년 9월 16일,
6. S. Du, M. Ibrahim, M. Shehata and W. Badawy, 게재확정일: 2020년 9월 22일

Journal of KSDT Vol. 19, No. 3, 2020