Overtaking Vehicle Detection Techniques based on Optical Flow and

Convolutional Neural Network

Lu-Ting Wu

and Huei-Yung Lin

Department of Electrical Engineering, National Chung Cheng University, 168 University Rd., Chiayi 621, Taiwan

Department of Electrical Engineering, Advanced Institute of Manufacturing with High-Tech Innovation,

National Chung Cheng University, 168 University Rd., Chiayi 621, Taiwan

Keywords:

Overtaking Vehicle Detection, ADAS, CNN.

Abstract:

As the rise of the intelligent vehicle applications in recent years, the development of onboard vision systems

for advanced driving assistance has become a popular research topic. This paper presents a real-time system

using a monocular camera mounted on the rear of a vehicle to perform overtaking detection for safe lane

change operations. In this work, the possible overtaking vehicle is ﬁrst located based on motion cues. The

candidate is then identiﬁed using Convolutional Neural Network (CNN) and tracked for behavior analysis

in a short period of time. We also propose an algorithm to solve the issue of repetitive patterns which is

commonly appeared in the highway driving. A series of experiments are carried out with real scene video

sequences recorded by a dashcam. The performance evaluation has demonstrated the effectiveness of the

proposed technique.

1 INTRODUCTION

In the past few decades, advanced driving assistance

systems (ADAS) have achieved great advances. They

are commonly used to improve the driving safety and

trafﬁc efﬁciency. Recently, the researches on advan-

ced driver assistance systems have gained a great mo-

mentum in vehicle safety issues. For examples, lane

departure warning system (LDWS), forward collision

warning system (FCWS), and lane change assistance

(LCA) are some typical applications of ADAS. At the

present, many high end vehicles are equipped with

ADAS directly from the original manufacturer. Some

trafﬁc recorders in the after market such as Mobileye

660 (Mobileye, 2017) and Papago P3 (Papago, 2017)

are also equipped with various ADAS functions. The

driving recorder or dashcam not only records the traf-

ﬁc conditions but also assists the driver to acquire the

trafﬁc information (Dai et al., 2017). The main ob-

jective is to provide a safe and comfort driving expe-

rience.

There are many kinds of sensors used in advan-

ced driving systems such as radar, lidar, and visual

sensor, etc. Among them, the imaging technology

using visual sensor has the greatest progress potential

in recent years. The visual sensors become cheaper

and lighter than ever before. They can provide rich

sensing information of the environment and therefore

be used to design more functions for driving assis-

tance. Some examples include vehicle identiﬁcation

(Liu et al., 2016), trafﬁc sign recognition (Luo et al.,

2017; Shi and Lin, 2017), front collision warning (Lin

et al., 2012), parking assistance (Fernandez-Llorca

et al., 2014), vehicle speed detection (Lin et al., 2008),

and other applications. In the past few years, image

processing speed is also greatly improved because of

the use of graphics processor unit (GPU). The pro-

gress of this hardware makes the image processing

task rapid and more suitable for advanced driving as-

sistance systems.

Vehicle drivers usually assess the surrounding

trafﬁc before changing lanes, and make turns after

checking their rear view and side mirrors. However,

even for those who follow the standard procedure, the

blind-zone of vehicles is still a source of danger and

often the cause of serious accidents. A blind-zone of a

vehicle can be described as “an area around the vehi-

cle that cannot be seen directly by the driver by look-

ing forwards or by using any of the vehicles standard

rear-view mirrors (internal and external) from the nor-

mal sitting position” (Hughes et al., 2009). According

to Taiwan Area National Freeway Bureau (MOTC,

2017), the main cause of major trafﬁc accidents in the

state roads was the result of “improper conversion la-

Wu, L. and Lin, H.

Overtaking Vehicle Detection Techniques based on Optical Flow and Convolutional Neural Network.

DOI: 10.5220/0006698001330140

In Proceedings of the 4th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2018), pages 133-140

ISBN: 978-989-758-293-6

133

Figure 1: The system ﬂowchart of the proposed overtaking detection technique. It consists of four basic modules: (1) pre-

processing and segmentation, (2) convolutional neural network, (3) repetitive pattern removal, and (4) tracking and behavior

detection.

nes” followed by “not paying attentions to the state in

front of the vehicle”.

In order to improve the driving safety, a dashcam

can be installed in the rear of the vehicle and used to

detect the blind-zone of the overtaking vehicles. It

is able to provide the necessary information for the

driver about the situation in advance, to avoid the

car accident due to an improper turn or lane change.

In this work, our experiments contain image sequen-

ces acquired from a variety of scenes during day and

night. To reduce the cost, a monocular camera is used

as the primary sensor. It is also a common approach

to analyze the trafﬁc condition using a static camera

system.

Because of the vehicle motion and the dynamic

change of the scene, the development of onboard ca-

mera systems is not a trivial task in terms of the

operation efﬁciency and result accuracy. The image

sequence is constantly blurred due to the rapid mo-

vement of the vehicle, which introduces unwanted ar-

tifacts when the overtaking vehicles approach the ca-

mera. Furthermore, the limited ﬁeld of view of the ca-

mera poses additional challenges on vehicle tracking.

The vehicle detection is generally more difﬁcult in the

urban areas and at night. In the urban areas, the en-

vironment is complicated and the vehicles will have

various kinds of behavior to respond to. At night, in

additional to the complex environment, the image se-

quences also contain less information due to the il-

lumination condition. In most situations, the feature

used for the detection of overtaking vehicles is only

the headlights.

2 RELATED WORK

There are many general vehicle detection techniques

in the literature. For related work, we mainly sur-

vey the vision based approaches for vehicle overta-

king and blind-zone detection. The camera system for

image acquisition is commonly installed in the front,

rear, or sides of the vehicle. For the camera mounted

in the front, the overtaking detection is based on the

vehicle’s motion cues. The advantage of using mo-

tion cues is the operation speed. It can still lead to

good results even when the vehicle is only partially

visible due to occlusion or limited ﬁeld of view of the

camera. Thus, a more complete and continuous tra-

jectory can be obtained. It is, however, easy to be

affected by noise and generates erroneous results.

In previous work, Hultqvistet et al. (Hultqvist

et al., 2014) and Chen et al. (Chen and Wu, 2015)

proposed efﬁciency detection overtaking algorithms

using optical ﬂow. Their approaches have the came-

ras placed in the front of the vehicle. As a result, they

cannot notify the driver about the occurrence of over-

taking in advance. Alonso et al. (Alonso et al., 2008)

presented a blind-zone overtaking vehicle detection

system using optical ﬂow with the cameras installed

below the side-view mirror. Wu et al. (Wu et al.,

2012) proposed an embedded blind-zone security sy-

stem with a camera mounted below the side-view mir-

ror. First, they detect low-features such as shadows,

edges and headlights to locate the vehicle, followed

by vehicle tracking and behavior analysis. The came-

ras below the side-view mirror can detect overtaking

vehicles better than others in the blind-zone. Howe-

ver, it requires a pair of cameras which are installed

in per external rear mirror and the functions of the ca-

meras are limited than other approaches.

There also exist some techniques with a camera

mounted in the rear of the vehicle to detect the over-

taking in the blind-zone. It is not only used to inform

the driver in advance, but also used by other advan-

ced driving assistance functions. For example, the

parking collision avoidance system can use the rear

camera to detect the obstacles in the back to avoid

collision. Dooley et al. (Dooley et al., 2016) installed

VEHITS 2018 - 4th International Conference on Vehicle Technology and Intelligent Transport Systems

134

a ﬁsheye camera in the rear of the vehicle, and used

the AdaBoost classiﬁcation technique and two wheel

detection methods to identify the blind-zone vehicles.

The vehicles were then tracked by the optical ﬂow

technique. Ramirez et al. (Ramirez et al., 2014) in-

stalled cameras in the front and the rear of the vehi-

cle. They used Deformable Parts Model (DPM) (Pan-

dey and Lazebnik, 2011) to combine optical ﬂow with

the mobile information to detect the overtaking vehi-

cles. According to their experimental results, combi-

ning these two methods they were able to increase the

accuracy of the detection compared to the use of an

appearance detector.

3 PROPOSED METHOD

In this paper, we propose a real-time system that uses

a monocular camera mounted in the rear of a vehicle

to detect overtaking and assist the driver to change la-

nes. The monocular camera is installed at the height

of about 1 meter from the ground. In our main met-

hod the architecture contains (1) pre-processing and

segmentation, (2) convolutional neural network, (3)

repetitive pattern removal, and (4) tracking and beha-

vior detection, as illustrated in Figure1.

3.1 Pre-Processing and Segmentation

In order to increase the speed of operations, we usu-

ally use only about half size of the original image by

removing unnecessary regions such as the sky, buil-

dings or trafﬁc signs located in the upper part of the

image. Thus, the region of interest (ROI) is ﬁrst de-

ﬁned before performing the image segmentation al-

gorithm. To obtain the motion clues in an image,

the settings on the tracking points are very important.

Chen et al. (Chen and Wu, 2015) use Canny edge de-

tection to extract the edges in the image, and then use

the optical ﬂow to calculate the edges of the motion

clues. Ramirezet et al. (Ramirez et al., 2014) detect

the strong corners, and then use the optical ﬂow to

calculate the corners of the motion clues. Although

using edges or corners as tracking features can derive

more robust results, it may increase the amount of ex-

tra computations. Further, it is hard for us to deter-

mine the number of tracking feature when the image

is complicated, which will also cause the uncertainty

of computing time.

To avoid the above problems, we set a ﬁxed num-

ber of tracking points in a ﬁxed location. First, we set

a tracking point per 10× 10 pixels in an ROI. Second,

we use the pyramid model of Lucas-Kanade optical

ﬂow (Lucas and Kanade, 1981) to calculate the op-

Figure 2: Directional classiﬁcation diagram. The yellow

lines represent the land marking. The blue lines illustrate

the objects (vehicles) approaching from the sides. The pink

lines illustrate the objects (vehicles) moving further away.

Figure 3: The optical ﬂow in the x-direction.

tical ﬂow information of the tracking points. Third,

according to the direction and measurement of the

tracking output, we divide it into ﬁve different cate-

gories: (i) road and sky, (ii) land marking, (iii) over-

taking vehicle, (iv) object further away, (v) uncertain

region.

When a vehicle is moving forward, there will be a

large amount of optical ﬂow around the vehicle. Ho-

wever, the road and sky are relatively smooth, and the

optical ﬂow is small. Therefore, if the optical ﬂow of

a feature point is very small, we can assign the point

to the road and sky region. If the optical ﬂow in the

y-direction is very large and moves to the vanishing

point, then the point is assigned to the land marking.

The points are considered as the overtaking vehicle

if the optical ﬂow in the x-direction is large and mo-

ves far away from the vanishing point. Otherwise,

if the optical ﬂow in the x-direction is large and mo-

ves towards the vanishing point, it indicates that some

object is moving further away. These conditions are

illustrated in Figure 2. We use the uncertain region to

conclude the situations not in one of the above crite-

ria.

Here we only use the x-direction of the optical

ﬂow to distinguish the overtaking vehicle from the ob-

ject moving further away, and discard the y-direction

of the optical ﬂow. The main reason is that the y-

direction of the optical ﬂow in two adjacent frames is

very small, and the approaching vehicle will be de-

formed (see Figure3). Thus, it is difﬁcult to use the

y-direction of optical ﬂow to distinguish the direction

of the object motion.

Overtaking vehicles are continually approaching,

Overtaking Vehicle Detection Techniques based on Optical Flow and Convolutional Neural Network

135

(a) (b) (c)

Figure 4: The accumulation of tracked points and image segmentation used to detect the approaching vehicles.

so the space correlation between two vehicles is used

to ﬁlter out the noise. We accumulate every tracking

point of the approaching object to a grayscale image

and binarize it. Then, a 10 × 10 mask is used to dilate

each tracking point, to make a connection between the

points, and ﬁlter out the small connection areas. An

illustration of these rules is shown in Figure 4(a). The

top is the original image, the middle is the result of

a single frame, and the bottom is the binarized result

from multiple frames.

When the vehicle makes turns or there are repe-

titive patterns in the image, the segmentation results

might not be correct, as illustrated in Figures 4(b) and

4(c) respectively. In the urban area, there are a va-

riety of curved roads. This makes the vehicle move

in a more complicated way, so the optical ﬂow of

the rear objects and background will be away from

the vanishing point. Another common problem is the

repetitive patterns in the images. When a vehicle is

in fast motion, another image point close to the ori-

ginal image point might be identiﬁed as the original

image point. With this continuous appearance, a re-

petitive pattern will be generated and false targets will

be tracked by the optical ﬂow algorithm. Since the re-

petitive patterns usually appear for a period of time in

the same image region, it is difﬁcult to ﬁlter out the

wrong features using the spatio-temporal correlation

of the image pixels. Figure 5 shows some repetitive

patterns commonly seen in Taiwan’s highway and lo-

cal road. To deal with this problem, we use the convo-

lutional neural network to identify the segmented area

and remove the non-vehicle objects to avoid the false

segmentation to produce false positives.

3.2 Convolutional Neural Network

The convolutional neural network architecture used in

this work is CaffeNet (BLVC, 2017). It is a replica-

tion of the model described in the AlexNet publica-

Figure 5: Common repetitive patterns in Taiwan.

tion (Krizhevsky et al., 2012) but with some differen-

ces. This network was originally designed to clas-

sify 1000 different categories of objects in the Ima-

geNet database. We performed ﬁne-tuning using our

data in BVLC CaffeNet model and changed the out-

put to 6 categories. It contains the front of the car,

the rear of the car, the motorcycle, repetitive patterns,

background and lane marks. Some of the images are

shown in Figure 6. The images are segmented by the

algorithm described in Section 3.1. The deep learning

framework, Caffe (Jia et al., 2014), is used to train

and evaluate the neural network. We added CaffeNet

to our system to identify the segmented area and re-

move the non-vehicle objects.

3.3 Repetitive Pattern Removal

In some situations, wrong segmentation, especially

when the algorithm is confused with a vehicle on the

opposite lane. As a result, the segmentation and CNN

identiﬁcation are erroneous at the same time. This

kind of error is due to the feature points of the opti-

cal ﬂow tracking on the repetitive patterns, but CNN

recognizes the vehicle in the opposite lane. One such

example is shown in Figure 7. To solve this problem,

we consider the case that a moving vehicle is in the

opposite lane and the repeated pattern occurs due to

the separation poles. As shown in Figure 8, they make

the optical ﬂow disorder, resulting in forward and bac-

kward optical ﬂow inconsistency.

VEHITS 2018 - 4th International Conference on Vehicle Technology and Intelligent Transport Systems

136

Figure 6: Some sample images in the training data used in this work.

Figure 7: The use of segmentation and CNN at the same

time can still be erroneous for repetitive patterns.

Figure 8: An illustration of messy optical ﬂow when repeti-

tive patterns occur.

We use multiple feature points to determine whet-

her the segmented region contains repetitive patterns.

First, “good features to track” is used to detect the fea-

ture point P(t) on the frame at time t (Shi and Tomasi,

1994). Then, its succeeding P(t + 1) is found by re-

ferring to the forward optical ﬂow f

= (u

, v

) from

the frame at t to the frame at t + 1. That is,

P(t + 1) = P(t) + f

(P(t))

Similarly, generated from the backward tracking,

the point P

(t) is related by the backward motion,

(t) = P

(t + 1) + f

−

t+1

(t + 1))

where f (t +1)

−

= (u

−

, v

−

)

(

t +1) is the backward op-

tical ﬂow from the frame at t + 1 to the frame at t.

Ideally, if the feature point P(t) does not belong

to a repetitive pattern and the optical ﬂow is correctly

estimated, then we have

P(t) − P

(t) = 0

However, from our experience, there are some errors

in the real situations, i.e.,

P(t) − P

(t) ≤ ε

where ε is the error between P(t) and P

(t), and is

very small. If most of the feature points on the target

are within ε, we will classify this object as a vehicle.

Otherwise, it is classiﬁed as a repetitive pattern and

thus eliminated.

3.4 Tracking and Behavior Detection

After the segmentation using CNN and the repetitive

pattern removal, we track the object rather until it di-

sappeared in the image or no overtaking. Continue

the previous step of “repetitive pattern removal” and

the detection of feature points, we use Lucas-Kanade

optical ﬂow for continuously tracking in order to

• detect the movement of objects for a period of

time,

Overtaking Vehicle Detection Techniques based on Optical Flow and Convolutional Neural Network

137

Figure 9: An example result of the overtaking vehicle de-

tection and tracking.

• overcome the shape, size and scale of the object

changes as it approaches the camera,

• get the operation speed quickly,

• have a more complete trajectory which can be pro-

vided at the edge or part of the ﬁeld.

A tracking result is shown in Figure9.

We expect that CNN is able to identify and remove

the repetitive patterns, and have the correct overtaking

detection. However, the camera resolution is too low,

and the image segmentation is not good enough for

identiﬁcation due to some errors. To further reduce

the false detection, we assess whether the direction of

tracking has continued to stay away from the vanis-

hing point for a period of time to help us effectively

remove the wrong detection.

3.5 Night Time

In general, the appearance of the vehicle at night can-

not be clearly obtained from the camera because the

brightness of the image is too low. The camera usu-

ally captures the information of the headlights. If only

the brightness is used to grab the headlights, it can be

easily disturbed by other background lights. Further-

more, the headlights of the vehicles in the sides do not

have the symmetry property, which makes the overta-

king detection at night a difﬁcult issue.

We present a method that uses motion cues to

combine the brightness information to capture the

overtaking vehicles. The steps of overtaking detection

at night includes: (1) pre-processing and segmenta-

tion, (2) brightness recognition, (3) tracking, and (4)

behavior detection. First, we use ROIs to remove the

image regions which are not important. Second, the

headlight areas are segmented when they approach

the vehicle by the algorithm described in Section 3.1.

Figure 10 shows a typical result. Third, the low brig-

htness areas are removed by binarization. Finally, we

track the brightness region and analyze its behavior

by the method given in Section 3.4. If it continues to

keep away from the vanishing point, it is considered

as the headlights of an overtaking vehicle.

4 IMPLEMENTATION AND

EXPERIMENTAL RESULTS

To evaluate the performance of our overtaking de-

tection method in the real world scenarios, we use

a dashcam to capture the dynamic images. Our

dashcam is installed in the rear of the vehicle at about

1 meter about ground. The original image size is

1200 × 800, and the sub-image region for process

is 600 × 200. We test our algorithms on a PC with

ubuntu 16.04 operating system, 4.0 GHz Intel CPU,

and ASUS GTX 1080 GPU. The execution time is di-

vided into three parts tabulated in Table 1.

There are totally 7,587 images collected in our

training data. The image sequences are segmented

using the algorithm described in Section 3.1 and mar-

ked manually for each image. Similarly, the vali-

dation data of 2,770 images are collected from 80-

minute video sequence and marked manually. The

recognition results for each category are tabulated in

Table 2. In our vehicle recognition module, it does

not matter if overtaking is identiﬁed or not since a

subsequent tracking stage will be performed. Howe-

ver, a serious false alarm will occur if the background

is recognized as overtaking. This usually happens

when making turns with a vehicle in the background

as shown in Figure 11.

It is not easy to evaluate the accuracy and efﬁ-

ciency of the real image sequence at the present time.

In particular, the algorithms for detecting vehicles by

motion cues do not have good benchmarks for perfor-

mance evaluation. There are totally 50 image sequen-

ces collected in our database. It contains 20 highway

image sequences, 15 city image sequences and 15

night image sequences. Each image sequence is about

two minutes. If we detect the vehicle before it is di-

sappeared in the image sequence, the overtaking de-

tection is successful. Otherwise, it is called missed.

It is considered as failed if the approaching vehicle is

not detected. The evaluation is tabulated in Table 3

and some result images are shown in Figure 12.

VEHITS 2018 - 4th International Conference on Vehicle Technology and Intelligent Transport Systems

138

Figure 10: Overtaking vehicle detection and tracking at night using the headlights.

Table 1: The execution time for individual stages of the algorithm.

Segmentation CNN Repetitive Patterns Removal Entire Time

time 6ms 3ms 1ms 6-10ms

Table 2: The recognition results for each category and the recall.

Background Lane mark Locomotive Repetitive pattern Front of vehicle Back of vehicle

Groundtruth 375 252 545 408 480 710

Background 322 1 2 0 1 11

Land mark 2 249 0 2 0 0

Locomotive 0 0 524 1 0 0

Repetitive pattern 25 2 0 405 0 3

Front of vehicle 13 0 16 0 470 34

Rear of vehicle 13 0 3 0 9 662

Recall 0.858 0.988 0.961 0.992 0.979 0.932

Table 3: The experimental results and performance evaluation.

Scene True overtakes Detected Missed False

City 102 102 0 13

Highway 79 79 0 1

At night 42 37 5 4

Figure 11: The false alarm might happen if the vehicle ma-

kes turns with another behind.

5 CONCLUSIONS

This paper presents a real-time system to detect over-

taking and assist in driving lane change based on vi-

sual cues from a dashcam . We use the motion cues

and combine with CNN to detect the vehicle appea-

rance. Compared to the low-order features and weak

classiﬁer it is faster and more robust. It is more sui-

ted for light and complex environments. In addition

to improving the limitation and shortcomings of the

existing methods, the proposed technique can main-

tain the operational efﬁciency and provide more com-

plete overtaking information. The performance eva-

luation has demonstrated the effectiveness of the pro-

posed techniques.

Figure 12: Some overtaking vehicle detection results. High

trafﬁc, local trafﬁc and night are shown in the top, middle,

and bottom, respectively.

ACKNOWLEDGMENTS

The support of this work in part by Create Electronic

Optical Co., LTD, Taiwan and the Ministry of Science

Overtaking Vehicle Detection Techniques based on Optical Flow and Convolutional Neural Network

139

and Technology of Taiwan under Grant MOST 104-

2221-E-194- 058-MY2, is gratefully acknowledged.

REFERENCES

Alonso, J. D., Vidal, E. R., Rotter, A., and Muhlenberg,

M. (2008). Lane-change decision aid system based on

motion-driven vehicle tracking. IEEE Transactions on

Vehicular Technology, 57(5):2736–2746.

BLVC (2017). CaffeNet. https://github.com/BVLC/caffe/

tree/master/models/bvlc reference caffenet.

Chen, Y. and Wu, Q. (2015). Moving vehicle detection

based on optical ﬂow estimation of edge. In 2015

11th International Conference on Natural Computa-

tion (ICNC), pages 754–758.

Dai, J. M., Liu, T. A. J., and Lin, H. Y. (2017). Road sur-

face detection and recognition for route recommenda-

tion. In 2017 IEEE Intelligent Vehicles Symposium

(IV), pages 121–126.

Dooley, D., McGinley, B., Hughes, C., Kilmartin, L., Jo-

nes, E., and Glavin, M. (2016). A blind-zone de-

tection method using a rear-mounted ﬁsheye camera

with combination of vehicle detection methods. IEEE

Transactions on Intelligent Transportation Systems,

17(1):264–278.

Fernandez-Llorca, D., Garcia-Daza, I., Martinez-Hellin, A.,

Alvarez-Pardo, S., and Sotelo, M. A. (2014). Par-

king assistance system for leaving perpendicular par-

king lots: Experiments in daytime/nighttime conditi-

ons. IEEE Intelligent Transportation Systems Maga-

zine, 6(2):57–68.

Hughes, C., Glavin, M., Jones, E., and Denny, P. (2009).

Wide-angle camera technology for automotive appli-

cations: a review. IET Intelligent Transport Systems,

3(1):19–31.

Hultqvist, D., Roll, J., Svensson, F., Dahlin, J., and Schn,

T. B. (2014). Detecting and positioning overtaking

vehicles using 1d optical ﬂow. In 2014 IEEE Intel-

ligent Vehicles Symposium Proceedings, pages 861–

866.

Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J.,

Girshick, R., Guadarrama, S., and Darrell, T. (2014).

Caffe: Convolutional architecture for fast feature em-

bedding. In Proceedings of the 22Nd ACM Internati-

onal Conference on Multimedia, MM ’14, pages 675–

678, New York, NY, USA. ACM.

Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012).

Imagenet classiﬁcation with deep convolutional neu-

ral networks. In Proceedings of the 25th Internatio-

nal Conference on Neural Information Processing Sy-

stems, NIPS’12, pages 1097–1105, USA. Curran As-

sociates Inc.

Lin, H. Y., Chen, L. Q., Lin, Y. H., and Yu, M. S. (2012).

Lane departure and front collision warning using a

single camera. In 2012 International Symposium

on Intelligent Signal Processing and Communications

Systems, pages 64–69.

Lin, H.-Y., Li, K.-J., and Chang, C.-H. (2008). Vehicle

speed detection from a single motion blurred image.

Image Vision Comput., 26(10):1327–1337.

Liu, X., Liu, W., Mei, T., and Ma, H. (2016). A deep

learning-based approach to progressive vehicle re-

identiﬁcation for urban surveillance. In Leibe, B.,

Matas, J., Sebe, N., and Welling, M., editors, Compu-

ter Vision - ECCV 2016 - 14th European Conference,

Proceedings, Part II, volume 9906 of Lecture Notes in

Computer Science, pages 869–884. Springer.

Lucas, B. D. and Kanade, T. (1981). An iterative image re-

gistration technique with an application to stereo vi-

sion. In Proceedings of the 7th International Joint

Conference on Artiﬁcial Intelligence - Volume 2, IJ-

CAI’81, pages 674–679, San Francisco, CA, USA.

Morgan Kaufmann Publishers Inc.

Luo, H., Yang, Y., Tong, B., Wu, F., and Fan, B. (2017).

Trafﬁc sign recognition using a multi-task convolutio-

nal neural network. IEEE Transactions on Intelligent

Transportation Systems, PP(99):1–12.

Mobileye (2017). Mobileye 660. http://

www.mobileye.com/.

MOTC (2017). Taiwan Area National Freeway Bureau,

MOTC. http://www.freeway.gov.tw/.

Pandey, M. and Lazebnik, S. (2011). Scene recognition and

weakly supervised object localization with deforma-

ble part-based models. In 2011 International Confe-

rence on Computer Vision, pages 1307–1314.

Papago (2017). Papago P3. http://www.papago.com.tw/.

Ramirez, A., Ohn-Bar, E., and Trivedi, M. (2014). Inte-

grating motion and appearance for overtaking vehicle

detection. In 2014 IEEE Intelligent Vehicles Sympo-

sium Proceedings, pages 96–101.

Shi, J. and Tomasi, C. (1994). Good features to track. In

1994 Proceedings of IEEE Conference on Computer

Vision and Pattern Recognition, pages 593–600.

Shi, J.-H. and Lin, H.-Y. (2017). A vision system for trafﬁc

sign detection and recognition. In 2017 IEEE Inter-

national Symposium on Industrial Electronics, pages

1596–1601.

Wu, B. F., Kao, C. C., Li, Y. F., and Tsai, M. Y. (2012).

A real-time embedded blind spot safety assistance sy-

stem. International Journal of Vehicular Technology,

2012:1–15.

VEHITS 2018 - 4th International Conference on Vehicle Technology and Intelligent Transport Systems

140