Available via license: CC BY 4.0
Content may be subject to copyright.
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access
Date of publication xxxx 00, 0000, date of current version June 19, 2019.
Digital Object Identifier
An Automatic Car Accident Detection
Method Based on Cooperative Vehicle
Infrastructure Systems
DAXIN TIAN1, (Senior Member, IEEE), CHUANG ZHANG1, XUTING DUAN1, AND XIXIAN
WANG1
1Beijing Advanced Innovation Center for Big Data and Brain Computing, Beijing Key Laboratory for Cooperative Vehicle Infrastructure Systems and Safety
Control, School of Transportation Science and Engineering, Beihang University, Beijing 100191, China.
Corresponding author: Xuting Duan (e-mail: duanxuting@buaa.edu.cn).
This research was supported in part by the National Natural Science Foundation of China under Grant Nos. 61672082 and 61822101,
Beijing Municipal Natural Science Foundation Nos. 4181002.
ABSTRACT Car accidents cause a large number of deaths and disabilities every day, a certain proportion
of which result from untimely treatment and secondary accidents. To some extent, automatic car accident
detection can shorten response time of rescue agencies and vehicles around accidents to improve rescue
efficiency and traffic safety level. In this paper, we proposed an automatic car accident detection method
based on Cooperative Vehicle Infrastructure Systems (CVIS) and machine vision. First of all, a novel image
dataset CAD-CVIS is established to improve accuracy of accident detection based on intelligent roadside
devices in CVIS. Especially, CAD-CVIS is consisted of various kinds of accident types, weather conditions
and accident location, which can improve self-adaptability of accident detection methods among different
traffic situations. Secondly, we develop a deep neural network model YOLO-CA based on CAD-CVIS and
deep learning algorithms to detect accident. In the model, we utilize Multi-Scale Feature Fusion (MSFF)
and loss function with dynamic weights to enhance performance of detecting small objects. Finally, our
experiment study evaluates performance of YOLO-CA for detecting car accidents, and the results show
that our proposed method can detect car accident in 0.0461 seconds (21.6FPS) with 90.02% average
precision (AP). In additionally, we compare YOLO-CA with other object detection models, and the results
demonstrate the comprehensive performance improvement on the accuracy and real-time over other models.
INDEX TERMS Car accident detection, CVIS, machine vision, deep learning
I. INTRODUCTION
ACCORDING to the World Health Organization, there
are about 1.35 million deaths and 20-50 million in-
juries as a result of the car accident globally every year
[1]. Especially, a certain proportion of deaths and injuries
are due to untimely treatment and secondary accidents [2],
which results from that rescue agency and vehicles around
accident cannot obtain quick response about the accident [3],
[4]. Therefore, it is vital important to develop an efficient
accident detection method, which can significantly reduce
both the number of deaths and injuries as well as the impact
and severity of accidents [5]. Under this background, many
fundamental projects and studies to develop efficient detec-
tion method have been launched for developing and testing
[6]–[10].
The traditional methods utilize vehicle motion parameters
captured by vehicular GPS devices to detect car accident,
such as acceleration and velocity. However, these methods
based on single type of features cannot meet the performance
need of accident detection in the aspect of accuracy and real-
time. With the development of computer and communication
technologies, Cooperative Vehicle Infrastructure System and
Internet of Vehicles have been developed rapidly in recent
years [11]–[13]. Moreover, the image recognition based on
video captured by intelligent roadside devices in CVIS has
become one of research hotspots in the field of intelligent
transportation system [14], [15]. For traffic situation aware-
ness, image recognition technology has advantages of high
efficiency, flexible installation and low maintenance costs.
Therefore, the image recognition has been applied to detec-
tion pedestrian, vehicle, traffic sign and so on successfully
[16]–[20]. In generally, there are many distinctive image and
VOLUME 4, 2016 1
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
video features in traffic accidents, such as vehicle collision,
rollover and so on. To some extent, these features can be
used to detect or predict car accidents. Accordingly, some
researchers apply the machine vision technology based on
deep-learning into methods of car accident detection. These
methods extract and process complex image features instead
of single vehicle motion parameter, which improves the
accuracy of detecting car accidents. However, the datasets
of these methods are mostly captured by car cameras or
cell phones of pedestrian, which is not suitable for roadside
devices in CVIS. In additionally, the reliability and real-time
performance of these methods need to be improved to meet
the requirements of car accident detection.
In this paper, we propose a data-driven car accident detec-
tion method based on CVIS, whose goal is improving effi-
ciency and accuracy of car accident response. With the goal,
we focus on such a general application scenario when there
is an accident on the road, roadside intelligent devices recog-
nize and locate it efficiently. First, we build a novel dataset,
Car Accident Detection for Cooperative Vehicle Infrastruc-
ture System dataset (CAD-CVIS), which is more suitable for
car accident detection based on roadside intelligent devices
in CVIS. Then, a deep learning model YOLO-CA based on
CAD-CVIS is developed to detect car accident. Especially,
we optimize the network of traditional deep learning models
YOLO [21] to build network of YOLO-CA, which is more
accurate and fast in detecting car accident. In additionally,
considering of wide shooting scope of roadside cameras in
CVIS, multi-scale feature fusion method and loss function
with dynamic weights are utilized to improve performance
of detecting small objects.
The rest of this paper is organized as follows: Section 2
gives an overviews of related work. We present the details of
our proposed method in Section 3. The performance evalua-
tion is discussed in Section 4. Finally, Section 5 conclude this
paper.
II. RELATED WORK
The car accident detection and notification method is a
challenging issue and has attracted a lot of attention from
researchers. They have proposed and applied various car ac-
cident detection methods. In generally, car accident detection
methods are mainly divided into the following two kinds:
vehicle running condition-based and accident video features-
based.
A. METHOD BASED ON VEHICLE RUNNING CONDITION
When an accident occurs, the motion state of the vehicle
will change dramatically. Therefore, many researchers pro-
posed the accident detection method by monitoring motion
parameters, such as acceleration, velocity and so on. [22]
used On Board Diagnosis (OBD) system to monitor speed
and engine status to detect a crash, and utilized smart-phone
to report the accident by Wi-Fi or cellular network. [23]
developed an accident detection and reporting system using
GPS, GPRS, and GSM. The speed of vehicle obtained from
High Sensitive GPS receiver is considered as the index for
detecting accidents, and the GSM/GPRS modem is utilized to
send the location of the accident. [24] presented a prototype
system called e-NOTIFY, which monitors the change of ac-
celeration to detect accident and utilize V2X communication
technologies to report it. To a certain extent, these methods
can detect and report car accidents in short time, and improve
the efficiency of car accidents warning. However, the vehi-
cle running condition before car accidents is complex and
unpredictable, and the accuracy of accident detection only
based on speed and acceleration may be low. In addition, they
rely too heavily on vehicular monitoring and communication
equipment, which may be unreliable or damaged in some
extreme circumstances, such as heavy canopy, underground
tunnel, and serious car accidents.
B. METHOD BASED VIDEO FEATURES
With the development of machine vision and artificial neural
network technology, more and more applications based on
video processing have been applied in transportation and
vehicle fields. Under this background, some researchers uti-
lized video features of the car accident to detect it. [25] pre-
sented a Dynamic-Spatial-Attention Recurrent Neural Net-
work (RNN) for anticipating accidents in dashcam videos,
which can predict accidents about 2 seconds before they
occur with 80% recall and 56.14% precision. [26] proposed
a car accident detection system based on first-person videos,
which detected anomalies by predicting the future locations
of car participants and then monitoring the prediction accu-
racy and consistency metrics. These methods also have some
limitations because of low penetration of vehicular intelligent
devices and shielding effects between vehicles.
There are also some other methods which use roadside
devices instead of vehicular equipments to obtain and pro-
cess video. [27] proposed a novel accident detection system
at intersection, which composed background images from
image sequence and detected accidents by using Hidden
Markov Model. [28] outlined a novel method for modeling
of interaction among multiple moving objects, and used the
Motion Interaction Field to detect and localize car accidents.
[29] proposed a novel approach for automatic road accident
detection, which was based on detecting damaged vehicles
from footage received from surveillance cameras installed
in roads. In this method, Histogram of gradients (HOG)
and Gray level co-occurrence matrix features were used to
train support vector machines. [30] presented a novel dataset
for car accidents analysis based on traffic Closed-Circuit
Television (CCTV) footage, and combined Faster Regions-
Convolutional Neural Network (R-CNN) and Context Min-
ing to detect and predict car accidents. The method in [30]
achieved 1.68 seconds in terms of Time-To-Accident mea-
sure with an Average Precision of 47.25%. [8] proposed a
novel framework for automatic car accident detection, which
learned feature representation from the spatio-temporal vol-
umes of raw pixel intensity instead of traditional hand-crafted
features. The experiments of method in [8] demonstrated it
2VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
can detect on average 77.5% accidents correctly with 22.5%
false alarms.
Compared with the methods based on vehicle running
condition, these methods improve the detection accuracy and
some of them even can predict accidents about 2 seconds be-
fore they occur. To some extent, these methods are significant
in decreasing the accident rate and improving traffic safety.
However, the detection accuracy of these methods is low and
the error rate is high, and the wrong accident information will
have a great impact on the normal traffic flow. Concerning
the core issue mentioned above, in order to avoid the draw-
backs of vehicular cameras, our proposed method utilizes the
roadside intelligent edge devices to obtain traffic video and
process image. Moreover, for sake of improving the accuracy
of accident detection method based on intelligent roadside
devices, we establish the CAD-CVIS dataset based on video
sharing websites, which is consisted of various kinds of
accident types, weather conditions and accident locations.
Moreover, we develop the model YOLO-CA to improve the
reliability and real-time performance among different traffic
conditions by combining deep learning algorithms and MSFF
method.
III. METHODS
A. METHOD OVERVIEW
Intelligent
roadside devices
Rescue agencies
Car accident!
CAD-CVIS
YOLO-CA
Data-driven
Accident?
Real-time
image
DSRC/5G
Y
FIGURE 1. The application scenario of the automatic car accident detection
method based on CVIS.
The Fig. 1 shows the application principle of our proposed
car accident detection method based CVIS. Firstly, the car ac-
cident detection application program with YOLO-CA model
is deployed on the edge server, which is developed based on
CAD-CVIS and deep learning algorithms. Then edge server
receives and processes the real-time image captured by road-
side cameras. Finally, the roadside communication unit will
broadcast the accident emergency messages to the relevant
vehicles and rescue agencies by DSRC and 5G networks. In
the rest of this section, we will present the details of CAD-
CVIS and YOLO-CA model.
B. CAD-CVIS
1) Data collection and annotation
There are two major challenges in collecting car accidents
data:(1) Access: access to roadside traffic cameras data is of-
ten limited. In addition, the accident data from transportation
administration is often not available for public uses because
of many legal reasons. (2) Abnormality: car accidents are rare
in the road compared with normal traffic conditions. In this
work, we try to draw support from video sharing websites to
search the videos and images including car accidents, such
as news report and documentary. In order to improve the
applicability of our proposed method to roadside edge device,
we only pick out the videos and images captured from a
traffic CCTV footage.
FIGURE 2. Data collection and annotation for the CAD-CVIS dataset.
Through the above steps, we obtain 633 car accidents
scenes, 3255 accident key frames and 225206 normal frames.
Moreover, the car accident scene only occupies a small
part of each accident frame. We utilize LabelImg [31] to
annotate the location of the accident in each frame in detail
to enhance the accuracy of locating accident. The high ac-
curacy enables emergency message be sent to the vehicles
that are in the same direction as accident more efficiently
and decrease the impact to the vehicles that are in the
opposite direction. The whole steps of data collection and
annotation are shown in Fig. 2. The CAD-CVIS dataset is
made available for research use through https://github.com/
zzzzzzc/Car-accident-detection.
2) Statistics of the CAD-CVIS
Statistics of the CAD-CVIS dataset can be found in Fig. 3.It
can be found that the CAD-CVIS dataset includes various
types of car accidents, which can improve the adaptability of
our method to different conditions. According to the number
of vehicles in the accident, the CAD-CVIS dataset includes
323 Single Vehicle Accident frames, 2449 Double Vehi-
cle Accidents frames and 483 Multiple Vehicle Accidents
frames. Moreover, the CAD-CVIS dataset covers a variety
of weather conditions, such as 2769 accident frames under
sunny condition, 268 frames under foggy condition, 52 acci-
dent frames under rainy condition and 166 accident frames
under snowy condition. Besides, there are 2588 frames of
accidents in the daytime and 667 accident frames at night.
In addition, the CAD-CVIS dataset contains 2281 frames
of accidents occurring at the intersection, 596 frames in the
urban road, 189 frames in the expressway and 189 frames in
the highway.
Comparison between CAD-CVIS and related datasets can
be found in Table. 1. The A in Table. 1 responses that there
is annotation of car accident in the dataset. R responses that
the videos and frames captured from the roadside CCTV
VOLUME 4, 2016 3
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
Single Double Multiple
Number of vehicles in the accident
0
300
600
900
1200
1500
1800
2100
2400
Number of traffic accident frames
(a)
Sunny Foggy Rainy Snowy
Weather condition
0
300
600
900
1200
1500
1800
2100
2400
2700
3000
Number of traffic accident frames
(b)
Daytime Night
Time of the accident
0
300
600
900
1200
1500
1800
2100
2400
2700
3000
Number of traffic accident frames
(c)
Intersection Urban road Expressway Highway
Location of the accident
0
300
600
900
1200
1500
1800
2100
2400
Number of traffic accident frames
(d)
FIGURE 3. Number of accident frames in CAD-CVIS categorized by different
indexes. (a) Accident Type (b) Weather condition (c) Accident time (d) Accident
location
TABLE 1. Comparison between CAD-CVIS and related datasets
Dataset name Scenes Frames or Duration A R M
UCSD Ped2 77 1636 frames ×X×
CUHK Avenue 47 3820 frames × × X
DAD 620 2.4 hours X×X
CADP 1416 5.2 hours ×X X
CAD-CVIS 632 3255+225206 frames X X X
footage. M responses that there are multiple road conditions
in dataset. Compared with CUHK Avenue [32], UCSD Ped2
[33] and DAD [25], CAD-CVIS contains more car accident
scenes, which can improve the adaptability of model based
on CAD-CVIS. Moreover, the frames of CAD-CVIS are all
captured from roadside CCTV footage, which is more suit-
able for the accident detection methods based on intelligent
roadside devices in CVIS.
C. OUR PROPOSED DEEP NEURAL NETWORK MODEL
In the task of car accident detection, we must not only
judge whether there is a car accident in the image, but
also accurately locate the car accident. That’s because the
accurate location guarantees that the RSU can broadcast the
emergency message to the vehicles affected by the accident.
The classification and location algorithms can be divided
into two kinds:(1) Two stage model, such as R-CNN [34],
Fast R-CNN [35], Faster R-CNN [36] and Faster R-CNN
with FPN [37]. These algorithms utilize selective research
and Region Proposal Network (RPN) to select about 2000
proposal regions in the image, and then detection objects
by the features of these regions extracted by CNN. These
region-based models locate objects accurately, but extracting
proposals take a great deal of time. (2) One stage model,
such as YOLO [21](You Only Look Once) and SSD (Single
Shot MultiBox Detector) [38]. These algorithms implement
location and classification by one CNN, which can provide
end to end detection service. Because of eliminating the
process of selecting the proposal regions, these algorithms
are very fast and still has guaranteeing accuracy. Considering
that accident detection requires high real-time performance,
we design the deep neural network based on one-stage model
YOLO [21].
1) Network Design
YOLO utilizes its particular CNN to complete classification
and location of multiple objects in an image at one time. In
the training process of YOLO, each image is divided into
S×Sgrids. If the center of an object falls into a grid cell,
that grid cell is responsible for detecting that object [39].
This design can improve the detection speed dramatically
and the detection accuracy with reference to global features.
However, it also will cause serious detection error when there
are more than one objects in one grids. Roadside cameras
have a wide scope of shooting, the accident area may be
small in the image. Inspired of the multi-scale feature fusion
(MSFF) network, in order to improve the performance of
model to detect small objects, we utilize 24 layers to achieve
image upsampling and obtain two different dimensional out-
put tensors. This new car accident detection model is called
as YOLO-CA, and the network structure diagram of YOLO-
CA is shown as Fig. 4.
As shown in Fig. 4, YOLO-CA is composed of 228 neural
network layers, and the number of each kind of layer is
shown in Table. 2. These layers constitute many kinds of
basic components of YOLO-CA network, such as DBL and
ResN. The DBL is the minimum components of YOLO-
CA network, which is composed of Convolution layer, Batch
Normalization layer and Leaky ReLU layer. ResN consists of
Zero Padding layer, DBL and N Resblock_units [40], which
is designed to avoid neural network degradation caused by
increased depth. Ups in Fig. 4 is upsampling layer, which is
utilized to improve the performance of YOLO-CA to detect
small objects. Concat is concatenate layer, which is used to
concatenate the layer in Darknet-53 and upsampling layer.
TABLE 2. Composition of YOLO-CA Network
Layer name Number
Input 1
Convolution 65
Batch Normalization 65
Leaky ReLU 65
Zero Padding 5
Add 23
Upsampling 1
Concatenate 1
Total 228
2) Detection principle
Fig. 5 shows the detection principle of YOLO-CA, which
includes extracting feature map and predicting bounding box.
4VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
DBL Res1 Res2 Res8 Res8 Res4
DBL DBL Conv
DBL UpS
DBL DBL Conv
Input
416x416x3
Output1
13x13x18
Output2
26x26x18
Concat
Darknet-53
DBL*5
DBL*5
DBL Conv BN Leaky
ReLU
Darkmetconv2D_BN_Leaky
ResN Zero
Padding DBL
Resblock_body
Res
Unit
Res_Unit*N
Resblock_unit
Res
Unit DBL DBL Add
FIGURE 4. The network structure of YOLO-CA.
416x416x3
13x13 grid
26x26 grid
13x13x18
26x26x18
BBox1 BBox2 BBox3
1
23
2
3
1
BBox1 BBox2 BBox3
FIGURE 5. The detection principle of YOLO-CA.
As shown in Fig. 5, YOLO-CA divides the input image into
13×13 grid and 26 ×26 grid. The first grid is responsible for
detecting the large objects, whereas the second grid makes up
for the inaccuracy of small target detection in the first grid.
The feature extraction networks corresponding to these two
grids are different, but the detection models of the objects
is similar. For ease of presentation, we regard the first grid
as example to explain the training steps of YOLO-CA. The
center of car accident region falls into the grid cell (7,5), so
this cell is responsible for detecting this car accident in the
whole training process. Then the cell (7,5) will predict three
bounding boxes, and each boxes includes six parameters:
x, y, w, h, C S, p. The (x, y)is the center point of the bound-
ing box, and the (w, h)is the ratio of width and height of
the bounding box to the whole image. The CS is confidence
score of bounding box, which represents how confident the
model is that the bounding box contains an object and how
accurate it thinks the box is that it predicts. Lastly, each
bounding box will predict class probability of car accident
p.
After the training of a batch of images, the loss of model
will be calculated, which is utilized to adjust the weights of
parameters. In the calculation of loss, let the ground truth
of an object is x∗, y∗, w∗, h∗, C S∗, p∗.S×Sis the number
of cells in grid, and Bis the number of predicted bounding
boxes of each grid cell. For each grid cell, the P r(Objects)
equals 1 when the cell contains center of object, whereas it
equals 0 when there is not center of object in the cell. For each
image, the loss of YOLO-CA is divided into the following
four parts:
•Loss of (x, y), which is calculated by (1). Where BCL
is binary cross entropy loss function, and the areaT ure
is defined as w∗∗h∗.
Lossxy =
S×S
X
i=1
B
X
j=1
P r (Objects) (2 −areaT ureij )
∗[BC L (xij) + B CL (yij )]
BC L (xij) = x∗
ij log xij +1−x∗
ij log (1 −xij )
BC L (yij) = y∗
ij log yij +1−y∗
ij log (1 −yij )
(1)
•Loss of (w, h), which is calculated by (2). Where
SD is square difference function. Especially, the (2 −
areaT ureij )in (1) and (2) is utilized to increase the
error punishment of small objects. Because that the
same errors of x, y, w, h cause more serious impact on
the detection effect of small object than that of large
VOLUME 4, 2016 5
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
Intersection
Union
t
p
IoU =
FIGURE 6. The definition of IoU.
object.
Losswh =
S×S
X
i=1
B
X
j=1
P r (Objects) (2 −areaT ureij )
∗1
2[SD (wij ) + S D (hij )]
SD (wij ) = wij −w∗
ij 2
SD (hij ) = hij −h∗
ij 2
(2)
•Loss of CS , which is calculated by (3). The loss of C S
can be divided into two parts: the confidence loss of
foreground and confidence loss of background.
LossCS =
S×S
X
i=1
B
X
j=1
P r (Objects)∗B CL (CSij )
+ (1 −P r (Objects)) ∗BCL (C Sij )
BC L (CSij ) = CS∗
ij log CSij
+1−CS ∗
ij log (1 −CSij )
(3)
Where CS ∗is defined by (4). In additionally, the IoU t
p
is defined in Fig. 6, which equals the intersection over
union (IoU) between the predicted bounding box and the
ground truth.
CS ∗= Pr(Objects)∗I oU t
p(4)
•Loss of p, which is calculated by (5)
LossCS =
S×S
X
i=1
B
X
j=1
P r (Objects)B CL (pij )
BC L (pij) = p∗
ij log pij +1−p∗
ij log (1 −pij )
(5)
For each image in training set, the total loss is defined as (6).
Especially, because that the multi-scale feature fusion is used
in YOLO-CA, the loss is the sum of conditions under S= 13
and S= 26. In additionally, the loss of each batch of images
is defined as (7).
Loss_img =Lossxy +Losswh +LossCS +Lossp(6)
Loss =1
b
b
X
k=1
Loss_imgk(7)
where the bin (7) is the size of batch.
IV. EXPERIMENT
In this section, we evaluate our proposed model YOLO-CA
on the CAD-CVIS dataset. First, we give the training results
of YOLO-CA, which include the change process of several
performance indexes. Then, we show the results of some
comparative experiments between YOLO-CA and other de-
tection models. Finally, the visual results are demonstrated
among various types and scales of car accident objects.
A. IMPLEMENTATION DETAILS
We implement our model in TensorFlow [41] under the op-
erating system Ubuntu 18.04 and perform experiments on a
system with Nvidia Titan Xp GPU. We divide the CAD-CVIS
dataset into three parts: (1) Training set (80%), which is used
to train the parameter weight of network. (2) Validation set
(5%), which is utilized to adjust hyperparameters, such as
learning rate and drop out rate. (3) Test set (15%), which is
used to evaluate the performance of different algorithms for
detecting car accident. In additionally, each part of dataset
contains all types of accident in Fig. 3. The batch size is set to
64, and the models are trained for up to 30000 iterations. The
initial learning rate is set to 0.001, and updating with iteration
parameter of 0.1/10000 iterations. The SGD optimizer with a
momentum of 0.9 is utilized to adjust parameters of network.
Moreover, we use a weight decay of 0.0005 to prevent model
overfitting.
B. RESULTS AND ANALYSIS
1) Training results of YOLO-CA
Fig. 7 shows the training results of YOLO-CA, including
the changes of precision, recall, IoU and loss of each batch
in iteration process. In the training process of YOLO-CA,
we regard the prediction result with IoU over 0.5 and right
classification as true result, and other predictions are all false
results. As shown in Table. 3, the prediction results can be
divided into four parts: (1) TP: Truth Positive. (2) FP: False
Positive. (3) FN: False Negative. (4) TN: True Negative. The
precision is defined as precision =T P
T P +F P and recall is
defined as recall =T P
T P +F N .
TABLE 3. Distribution map of prediction results
Ground truth
Positive Negative
Prediction Positive TP FP
Result Negative FN TN
As shown in Fig. 7a, with the increasing of iterations, the
precision of YOLO-CA is increasing gradually and converge
over 90%. Moreover, recall eventually converges to more
than 95%. In terms of locating performance of YOLO-CA
6VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
0 0.5 1 1.5 2 2.5 3
Batches 104
0
0.2
0.4
0.6
0.8
1
Precision
(a)
0 0.5 1 1.5 2 2.5 3
Batches 104
0
0.2
0.4
0.6
0.8
1
Recall
(b)
0 0.5 1 1.5 2 2.5 3
Batches 104
0
0.2
0.4
0.6
0.8
1
IoU
(c)
0 0.5 1 1.5 2 2.5 3
Batches 104
0
0.4
0.8
1.2
1.6
2
Loss
(d)
FIGURE 7. The training results of YOLO-CA. (a) Precision (b) Recall (c) IoU
(d) Loss
in training set, IoU finally stabilizes above 0.8. The Fig. 7d
shows the decreasing process of loss of YOLO-CA in (7),
and the final convergence of loss is less than 0.2.
2) Comparative experiments and visual results
The comparative experiments are conducted for comparing
seven detection models: (1) One-stage models: SSD, our
proposed YOLO-CA, traditional YOLO-v3 and YOLO-v3
without MSFF (Multi-Scale Feature Fusion). (2) Two-stage
models: Fast R-CNN, Faster R-CNN and Faster R-CNN with
FPN. In order to comparatively demonstrate the validation
of YOLO-CA as well as confirm its strength in terms of the
comprehensive performance on the accuracy and real-time,
the following indexes are selected for comparison among the
seven models:
•Average Precision (AP) that is defined as the average
value of precision under different recall, which can be
changed by adjusting threshold of classification confi-
dence. AP index evaluate the accuracy performance of
detection models. The average precision can be calcu-
lated by (8).
AP =Z1
0
precision(r)(8)
where ris recall.
•Average Intersect over Union (Average IoU) that is
utilized to evaluate the object locating performance of
detection models. The Average IoU is the average value
of IoUs between every prediction bounding box and
corresponding ground truth.
•Frames Per Second (FPS). Inference time is defined as
the average time cost of detecting a frame among test
set. FPS is the reciprocal of inference time, which is
defined as the average number of frames that can be
detected in one second. This indexes evaluate real-time
performance of detection models.
0 0.2 0.4 0.6 0.8 1
Recall
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Precision
Fast R-CNN (AP=77.65%)
Faster R-CNN (AP=82.02%)
Faster R-CNN with FPN (AP=90.66%)
SSD (AP=83.40%)
YOLOv3 without MSFF (AP=83.16%)
YOLOv3 (AP=86.20%)
YOLO-CA (AP=90.02%)
(a)
0
0.2
0.4
0.6
0.8
1
Average IoU
Fast
R-CNN
Faster
R-CNN
Faster
R-CNN
with FPN
SSD YOLOv3
without
MSFF
YOLOv3 YOLO-CA
(b)
FIGURE 8. The AP and IoU results of different models. (a) Precision-Recall
curve (b) Average IoU
The Fig. 8a shows the Precision-Recall curve of detection
models among test set. It can be found that Faster R-CNN
with FPN and our proposed YOLO-CA have obvious advan-
tages in accuracy performance than the other models. In addi-
tionally, our-proposed YOLO-CA can achieve 90.02% of AP,
which is slightly lower than that of Faster R-CNN with FPN
(90.66%) and higher than those of Fast R-CNN (77.66%),
Faster R-CNN (82.02%), SSD (83.40%), YOLOv3 without
MSFF (83.16%) and YOLOv3 (86.20%). Average IoU is a
vital important index to evaluate locating performance of
detection models. Moreover, accurate location is critical to
car accident detection and notification, and higher locating
performance can improve the safety of the vehicles around
accident. As shown in Fig. 8b, YOLO-CA can achieve about
0.73 of Average IoU, which is lower than that of Faster R-
CNN with FPN (0.75) and higher than those of Fast R-CNN
(0.58), Faster R-CNN (0.66), SSD (0.69), YOLOv3 without
MSFF (0.65) and YOLOv3 (0.71).
VOLUME 4, 2016 7
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
TABLE 4. AP and IoU results of different models among different scales of object
AP(%) Average IoU
Method Large Medium Small Large Medium Small
Fast R-CNN 82.11 80.3 49.20 0.68 0.63 0.46
Faster R-CNN 89.66 86.79 56.79 0.71 0.69 0.51
Faster R-CNN with FPN 93.07 92.15 78.63 0.74 0.72 0.68
SSD 91.37 89.65 60.59 0.73 0.71 0.58
YOLOv3 without MSFF 90.18 90.40 58.89 0.72 0.68 0.50
YOLOv3 90.45 91.17 67.78 0.74 0.73 0.60
YOLO-CA 93.87 91.51 76.51 0.76 0.74 0.64
In order to compare and analysis the performance of mod-
els in details, the objects of test set is divided into three parts
according to different scales of objects:(1) Large: the area of
object is larger than one tenth of image size. (2) Medium:
the area of object is over the interval [1/100, 1/10] of image
size. (3) Small: the area of object is less than one-hundredth
of image size.
The Table. 4 shows the AP and IoU results of the seven
models among different scales of object. We can intuitively
see that the scales of objects significantly affect the accuracy
and locating performance of detection models. It can be
found that our proposed YOLO-CA has obvious advantages
in AP and Average IoU than Fast R-CNN, Faster R-CNN
and YOLOv3 without MSFF, especially among small scale
of objects. There is not MSFF process in the above three
models, which results in that they detection the objects only
rely on the top-level features. However, although there is
rich semantic information in top-level features, the location
information of objects is rough, which does not benefit to
locate the bounding box of objects correctly. On the contrary,
there is little semantic information in low-level features with
high resolution, but the location information of objects is
accurate. For small scales of objects, they make up a small
proportion of the whole frame, and their location information
is easily lost through multiple convolution processes. YOLO-
CA utilizes MSFF to combine top-level features and low-
level features, and then makes a prediction in each fused
feature layer. This process reserves the rich semantic in-
formation and accurate location information simultaneously,
so YOLO-CA has better performance in AP and Average
IoU than Fast R-CNN, Faster R-CNN and YOLOv3 with-
out MSFF. For SSD, it uses pyramidal feature hierarchy to
obtain multi-scale feature maps. But to avoid using low-
level features SSD foregoes reusing already computed layers
and instead builds the pyramid starting from high up in the
network and then by adding several new layers. So SSD
misses the opportunity to reuse the higher-resolution maps
of feature hierarchy, which are vital important for detecting
small objects [37]. Moreover, the performance of backbone
of YOLO-CA (Darknet53) is better than that of SSD (VGG-
16) because of using residual networks to avoid degradation
problem of deep neural network. Therefore, YOLO-CA can
achieve better results of AP and Average IoU than SSD.
Compared with YOLOv3, YOLO-CA utilizes loss function
with dynamic weights to balance the influence of location
loss among different scales of objects. This process increases
the error punishment of small objects, because that the
same errors of x, y, w, h cause more serious impact on the
detection effect of the small object than that of the large
object. Consequently, YOLO-CA has obvious advantages in
AP and Average IoU of small objects than YOLOv3. The
MSFF processes of Faster R-CNN with FPN and YOLO-
CA are similar, feature pyramid networks is used to extract
feature maps of different scales and fuse these maps to obtain
features with high-semantic and high-resolution. Faster R-
CNN utilizes RPN to select about 20000 proposal regions,
whereas there are only 13∗13∗3+26∗26∗3 = 2535 candidate
bounding boxes in YOLO-CA. This difference results in
Faster R-CNN has slight advantages in accuracy performance
than YOLO-CA, but also causes serious disadvantages in
real-time performance.
0
3
6
9
12
15
18
21
24
27
30
FPS
Fast
R-CNN
Faster
R-CNN
Faster
R-CNN
with FPN
SSD YOLOv3
without
MSFF
YOLOv3 YOLO-CA
FIGURE 9. The FPS of different models.
Fig. 9 shows the FPS results of different models among
test set. It can be found that the FPS of one-stage models
is obviously higher than that of two-stage models. This low
performance of the two-stage models results from a great deal
of time cost of selecting proposal regions.
Fast R-CNN utilizes time-consuming selective research al-
gorithm to select proposal regions based on color and texture
features, which results in that Fast R-CNN only achieves 0.4
of FPS. Faster R-CNN uses the RPN that share convolutional
layers with state-of-the-art object detection networks instead
of selective research to generate proposals. Benefiting from
8VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
RPN, Faster R-CNN achieve about 3.5 of FPS among test set
(Faster R-CNN:3.5, Faster R-CNN with FPN:3.6).
Although Faster R-CNN obtains significantly improve-
ment of real-time performance compared with Fast R-CNN,
there is still a big gap with one-stage models. That is be-
cause one-stage models abandon the process of selecting
proposal regions and utilize one CNN to implement location
and classification of objects. As shown in Fig. 9, SSD can
achieve 15.6 of FPS among test set. The other three models
based on YOLO utilize the backbone Darknet-53 instead of
VGG-16 in SSD, and computation of the former network
is significantly less than the latter because of using the
residual networks. Therefore, the real-time performance of
SSD is lower than YOLO-based models in our experiments.
In additionally, our proposed YOLO-CA simplifies the MSFF
networks of YOLOv3. So YOLO-CA can achieve 21.7 of
FPS, which is higher than that of YOLOv3 (about 19.1). Be-
cause of lacking MSFF process in YOLOv3 without MSFF,
it has better real-time performance (about 23.6 of FPS) than
YOLO-CA, but this lacking results in serious performance
penalties of AP.
Fig. 10 show some visual results of the seven models
among different scales of objects. It can be found that there
is a false positive in the large objects detection results of Fast
R-CNN, but the other six models all have high accuracy and
locating performance in large objects in Fig. 10. However,
the locating performance of Fast R-CNN, Faster R-CNN,
SSD, and YOLOv3 without MSFF decrease significantly in
medium object frame (1), and the prediction bounding box
cannot fitting out the contour of car accident. Moreover, Fast
R-CNN, SSD, and YOLO-without MSFF cannot detect the
car accident in small object frame (1). In additionally, except
for Faster R-CNN with FPN and YOLO-CA, other models
have serious location error in small object frame (3).
3) Comparison of comprehensive performance and
practicality
As analyzed above, it can be found that our proposed YOLO-
CA has performance advantages of detecting car accident
than Fast R-CNN, Faster R-CNN, SSD, and YOLOv3 in
terms of accuracy, locating and real-time performance. For
YOLOv3 without MSFF, the FPS of it (23.6) is higher than
that of YOLO-CA (21.7), and this difference is acceptable in
the practical application of detecting car accident. However,
the AP of YOLO-CA is significantly higher than that of
YOLOv3 without MSFF, especially for small scales of object
(76.51% vs 58.89%). Compared with Faster R-CNN with
FPN, YOLO-CA can approach the AP of it (90.66% vs
90.03%) with an obvious speed advantage. Faster R-CNN
cost about 277ms on average to detect one frame, whereas
YOLO-CA only need 46 ms, which illustrates the speed
of YOLO-CA is about 6x faster than Faster R-CNN with
FPN. Car accident detection in CVIS requires high real-time
performance because of the high dynamics of vehicles. To
summarize, our proposed YOLO-CA have higher practicality
and comprehensive performance on accuracy and real-time.
4) Comparison with other car accident detection methods
Although other car accident detection methods utilize a small
private collection of datasets and do not make them public
so comparing them may not be fair at this stage. But still,
we list the performance achieved by these methods on their
individual datasets. ARRS [3] achieve about 63% AP with
6% false alarms. The method of [42] achieve 89.50% AP.
DSA-RNN [25] achieve about 80% recall and 56.14% AP.
The method in [30] achieve about 47.25% AP. The method
of [8] achieve 77.5% AP and 22.5% false alarms. Moreover,
the number of accident scenes of the datasets utilized in these
methods is limited, which will result in poor adaptability for
new scenarios.
V. CONCLUSION
In this paper, we have proposed an automatic car accident
detection method based on CVIS. First of all, we present the
application principles of our proposed method in the CVIS.
Secondly, we build a novel image dataset CAD-CVIS, which
is more suitable for car accident detection method based on
intelligent roadside devices in CVIS. Then we develop the car
accident detection model YOLO-CA based on CAD-CVIS
and deep learning algorithms. In the model, we combine the
multi-scale feature fusion and loss function with dynamic
weights to improve real-time and accuracy of YOLO-CA.
Finally, we show the simulation experiments results of our
method, which demonstrates our proposed methods can de-
tect car accident in 0.0461 seconds with 90.02% AP. More-
over, the comparative experiments results show that YOLO-
CA has comprehensive performance advantages of detecting
car accident than other detection models, in terms of accuracy
and real-time.
REFERENCES
[1] WHO, “Global status report on road safety 2018,” https://www.who.int/
violence_injury_prevention/road_safety_status/2018/en/.
[2] H. L. Wang and M. A. Jia-Liang, “A design of smart car accident rescue
system combined with wechat platform,” Journal of Transportation Engi-
neering, 2017.
[3] Y. Ki and D. Lee, “A traffic accident recording and reporting model at
intersections,” IEEE Transactions on Intelligent Transportation Systems,
vol. 8, no. 2, pp. 188–194, June 2007.
[4] W. Hao and J. Daniel, “Motor vehicle driver injury severity study under
various traffic control at highway-rail grade crossings in the united states,”
Journal of Safety Research, vol. 51, pp. 41–48, 2014.
[5] J. White, C. Thompson, H. Turner, B. Dougherty, and D. C. Schmidt,
“Wreckwatch: Automatic traffic accident detection and notification with
smartphones,” Mobile Networks and Applications, vol. 16, no. 3, pp. 285–
303, 2011.
[6] S. Sadek, A. Al-Hamadi, B. Michaelis, and U. Sayed, “Real-time auto-
matic traffic accident recognition using hfg,” in International Conference
on Pattern Recognition, 2010.
[7] A. Shaik, N. Bowen, J. Bole, G. Kunzi, D. Bruce, A. Abdelgawad, and
K. Yelamarthi, “Smart car: An iot based accident detection system,” in
2018 IEEE Global Conference on Internet of Things (GCIoT). IEEE,
2018, pp. 1–5.
[8] D. Singh and C. K. Mohan, “Deep spatio-temporal representation for
detection of road accidents using stacked autoencoder,” IEEE Transactions
on Intelligent Transportation Systems, vol. 20, no. 3, pp. 879–887, March
2019.
[9] M. Zheng, T. Li, R. Zhu, J. Chen, Z. Ma, M. Tang, Z. Cui, and Z. Wang,
“Traffic accident’s severity prediction: A deep-learning approach-based
cnn network,” IEEE Access, vol. 7, pp. 39 897–39 910, 2019.
VOLUME 4, 2016 9
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
Small
Mediun
Large
Fast R-CNN YOLOv3
Without MSFF YOLOv3
Faster R-CNN Faster R-CNN
With FPN SSD YOLO-CA
(1)
(2)
(3)
(1)
(2)
(3)
(1)
(2)
(3)
FIGURE 10. Some visual results of the seven models among different scales of objects.
[10] L. Zheng, Z. Peng, J. Yan, and W. Han, “An online learning and unsuper-
vised traffic anomaly detection system,” Advanced Science Letters, vol. 7,
no. 1, pp. 449–455, 2012.
[11] F. Yang, S. Wang, J. Li, Z. Liu, and Q. Sun, “An overview of internet of
vehicles,” China Communications, vol. 11, no. 10, pp. 1–15, Oct 2014.
[12] C. Ma, W. Hao, A. Wang, and H. Zhao, “Developing a coordinated
signal control system for urban ring road under the vehicle-infrastructure
connected environment,” IEEE Access, vol. 6, pp. 52 471–52478, 2018.
[13] S. Zhang, J. Chen, F. Lyu, N. Cheng, W. Shi, and X. Shen, “Vehicular
communication networks in the automated driving era,” IEEE Communi-
cations Magazine, vol. 56, no. 9, pp. 26–32, 2018.
[14] Y. Wang, D. Zhang, Y. Liu, B. Dai, and L. H. Lee, “Enhancing transporta-
tion systems via deep learning: A survey,” Transportation research part C:
emerging technologies, 2018.
[15] G. Wu, F. Chen, X. Pan, M. Xu, and X. Zhu, “Using the visual intervention
influence of pavement markings for rutting mitigation–part i: preliminary
experiments and field tests,” International Journal of Pavement Engineer-
ing, vol. 20, no. 6, pp. 734–746, 2019.
[16] S. Ramos, S. Gehrig, P. Pinggera, U. Franke, and C. Rother, “Detecting
unexpected obstacles for self-driving cars: Fusing deep learning and
geometric modeling,” in 2017 IEEE Intelligent Vehicles Symposium (IV).
IEEE, 2017, pp. 1025–1032.
[17] T. Qu, Q. Zhang, and S. Sun, “Vehicle detection from high-resolution
aerial images using spatial pyramid pooling-based deep convolutional
neural networks,” Multimedia Tools and Applications, vol. 76, no. 20, pp.
21 651–21 663, 2017.
[18] D. Dooley, B. McGinley, C. Hughes, L. Kilmartin, E. Jones, and
M. Glavin, “A blind-zone detection method using a rear-mounted fisheye
camera with combination of vehicle detection methods,” IEEE Transac-
tions on Intelligent Transportation Systems, vol. 17, no. 1, pp. 264–278,
Jan 2016.
[19] X. Changzhen, W. Cong, M. Weixin, and S. Yanmei, “A traffic sign
detection algorithm based on deep convolutional neural network,” in 2016
IEEE International Conference on Signal and Image Processing (ICSIP),
Aug 2016, pp. 676–679.
[20] S. Zhang, C. Bauckhage, and A. B. Cremers, “Efficient pedestrian detec-
10 VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
tion via rectangular features based on a statistical shape model,” IEEE
Transactions on Intelligent Transportation Systems, vol. 16, no. 2, pp. 763–
775, April 2015.
[21] J. Redmon and A. Farhadi, “Yolov3: An incremental improvement,” arXiv:
Computer Vision and Pattern Recognition, 2018.
[22] J. Zaldivar, C. T. Calafate, J. C. Cano, and P. Manzoni, “Providing accident
detection in vehicular networks through OBD-II devices and android-
based smartphones,” Proceedings - Conference on Local Computer Net-
works, LCN, no. April 2014, pp. 813–819, 2011.
[23] M. S. Amin, J. Jalil, and M. B. I. Reaz, “Accident detection and reporting
system using gps, gprs and gsm technology,” in International Conference
on Informatics, 2012.
[24] M. Fogue, P. Garrido, F. J. Martinez, J. C. Cano, C. T. Calafate, and P. Man-
zoni, “Automatic accident detection: Assistance through communication
technologies and vehicles,” IEEE Vehicular Technology Magazine, vol. 7,
no. 3, pp. 90–100, 2012.
[25] F.-H. Chan, Y.-T. Chen, Y. Xiang, and M. Sun, “Anticipating accidents
in dashcam videos,” in Computer Vision – ACCV 2016. Springer
International Publishing, 2017, pp. 136–153.
[26] Y. Yao, M. Xu, Y. Wang, D. J. Crandall, and E. M. Atkins, “Unsupervised
traffic accident detection in first-person videos.” arXiv: Computer Vision
and Pattern Recognition, 2019.
[27] S. Kamijo, Y. Matsushita, K. Ikeuchi, and M. Sakauchi, “Traffic Mon-
itoring and Accident Detection at Intersections,” IEEE Transactions on
Intelligent Transportation Systems, vol. 1, no. 2, pp. 108–117, 2000.
[28] K. Yun, H. Jeong, K. M. Yi, S. W. Kim, and J. Y. Choi, “Motion interaction
field for accident detection in traffic surveillance video,” Proceedings -
International Conference on Pattern Recognition, pp. 3062–3067, 2014.
[29] V. Ravindran, L. Viswanathan, and S. Rangaswamy, “A Novel Approach to
Automatic Road-Accident Detection using Machine Vision Techniques,”
International Journal of Advanced Computer Science and Applications,
vol. 7, no. 11, pp. 235–242, 2016.
[30] A. Shah, J. Lamare, T. Nguyenanh, and A. G. Hauptmann, “Cadp: A novel
dataset for cctv traffic camera based accident analysis,” advanced video
and signal based surveillance, pp. 1–9, 2018.
[31] Tzutalin, “Labelimg,” Git code, 2015,
https://github.com/tzutalin/labelImg.
[32] C. Lu, J. Shi, and J. Jia, “Abnormal event detection at 150 FPS in MAT-
LAB,” Proceedings of the IEEE International Conference on Computer
Vision, pp. 2720–2727, 2013.
[33] V. Mahadevan, W. Li, V. Bhalodia, and N. Vasconcelos, “Anomaly
detection in crowded scenes,” 2010 IEEE Computer Society Conference
on Computer Vision and Pattern Recognition, pp. 1975–1981, 2010.
[Online]. Available: http://ieeexplore.ieee.org/document/5539872/
[34] R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich feature hierarchies
for accurate object detection and semantic segmentation,” in The IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), June
2014.
[35] R. Girshick, “Fast r-cnn,” Computer Science, 2015.
[36] S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: towards real-
time object detection with region proposal networks,” in International
Conference on Neural Information Processing Systems, 2015.
[37] T.-Y. Lin, P. Dollár, R. B. Girshick, K. He, B. Hariharan, and S. J. Belongie,
“Feature pyramid networks for object detection,” 2017 IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), pp. 936–944, 2016.
[38] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C. Y. Fu, and A. C.
Berg, “Ssd: Single shot multibox detector,” in European Conference on
Computer Vision, 2016.
[39] J. Redmon, S. K. Divvala, R. B. Girshick, and A. Farhadi, “You only look
once: Unified, real-time object detection,” computer vision and pattern
recognition, pp. 779–788, 2016.
[40] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition,” in 2016 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), June 2016, pp. 770–778.
[41] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S.
Corrado, A. Davis, J. Dean, M. Devin et al., “Tensorflow: Large-scale ma-
chine learning on heterogeneous distributed systems,” arXiv: Distributed,
Parallel, and Cluster Computing, 2015.
[42] K. Yun, H. Jeong, K. M. Yi, S. W. Kim, and J. Y. Choi, “Motion interaction
field for accident detection in traffic surveillance video,” in 2014 22nd
International Conference on Pattern Recognition, Aug 2014, pp. 3062–
3067.
VOLUME 4, 2016 11
