TRAFFIC LIGHTS DETECTION IN ADVERSE CONDITIONS

USING COLOR, SYMMETRY AND SPATIOTEMPORAL

INFORMATION

George Siogkas, Evangelos Skodras and Evangelos Dermatas

Electrical and Computer Engineering Department, University of Patras, Rio, Patras, Greece

Keywords: Traffic Light Detection, Computer Vision, Fast Radial Symmetry, Adverse Conditions, Spatiotemporal

Persistency.

Abstract: This paper proposes the use of a monocular video camera for traffic lights detection, in a variety of

conditions, including adverse weather and illumination. The system incorporates a color pre-processing

module to enhance the discrimination of red and green regions in the image and handle the “blooming

effect” that is often observed in such scenes. The fast radial symmetry transform is utilized for the detection

of traffic light candidates and finally false positive results are minimized using spatiotemporal persistency

verification. The system is qualitatively assessed in various conditions, including driving in the rain, at night

and in city roads with dense traffic, as well as their synergy. It is also quantitatively assessed on a publicly

available manually annotated database, scoring high detection rates.

1 INTRODUCTION

Advanced Driver Assistance Systems (ADAS) are

new emerging technologies, primarily developed as

automated advisory systems to enhance driving

safety. One of the most challenging problems for such

systems is driving in urban environments, where the

visual information flow is very dense and can cause

fatigue, or distract the driver. Aside from moving

obstacles like cars and pedestrians, road signs and

traffic lights (TL) significantly influences the

reactions and decisions of the driver, which have to be

made in real time. The inability to make correct

decisions can lead to serious, even fatal, accidents.

Traffic light violations in particular, are one of the

most common causes of road crashes worldwide.

This is where ADAS can provide help that may

prove life-saving. A very big portion of such

systems is based on visual information processing

using computer vision methods. Whether in the form

of completely automated driving systems, using only

visual information, or in the form of driver alert

systems, technology can provide crucial assistance

in the efforts to reduce car accidents. Both the

aforementioned systems have to include a reliable

and precise Traffic Light Detection (TLD) module,

so that accidents in intersections are mitigated.

The idea of using vision for ADAS in urban

environments so that TLD can be achieved was first

introduced in the late 90’s, by (Loy and Zelinsky,

2002) who proposed a computer vision based Stop

& Go algorithm using a color on-board camera.

However, the use of computer vision for ADAS

bloomed in the next decade, as computer processor

speeds reached a point that enabled real-time

implementation of complex algorithms. The work of

(Lindner et al., 2004) proposes the fusion of color

cameras, GPS and vehicle data to increase the

robustness of their TLD algorithm, which is

followed by a tracking and a classification module.

The TLD part uses RGB color values, texture and

shape features. In the HSV space, color thresholding

followed by a Gaussian filtering process and a

verification of TL candidates is the approach of (In-

Hak et al., 2006) to detect TLs in crossroads. A

similar approach is followed by (Yehu Shen et al.,

2009), based on HSV images and a Gaussian

distribution-based model acquired by training

images. A post processing phase utilizes shape and

temporal consistency information to enhance the

results. A more straight-forward process is proposed

by (M. Omachi and Omachi, 2009), who locate

candidate regions in the normalized RGB color

space and validate the results using edge and

620

Siogkas G., Skodras E. and Dermatas E..

TRAFFIC LIGHTS DETECTION IN ADVERSE CONDITIONS USING COLOR, SYMMETRY AND SPATIOTEMPORAL INFORMATION.

DOI: 10.5220/0003855806200627

In Proceedings of the International Conference on Computer Vision Theory and Applications (VISAPP-2012), pages 620-627

ISBN: 978-989-8565-03-7

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

symmetry detection. Color information has been

ignored by (de Charette and Nashashibi, 2009a),

who use grayscale spot detection followed by an

Adaptive Template Matcher, achieving high

precision and recall rates in real time. Their results

were tested thoroughly and compared to the results

of an AdaBoost method using manually annotated

videos in (de Charette and Nashashibi, 2009b). The

problem of TLD in day and night scenes has been

addressed by (Chunhe Yu et al., 2010) using RGB

thresholding and shape dimensions to detect and

classify lights in both conditions.

By reviewing related literature to date, the

following conclusions can be drawn about vision

based TL detection:

 Researchers have not always used color

information; when they do, they prefer either

the HSV, or RGB color spaces.

 Many groups have used empirical thresholds

that are not optimal for all possible driving

conditions (i.e. shadows, rain, and night).

Generally, adverse conditions are not very

frequently addressed.

 Symmetry is frequently used, either with

novel detection techniques, or with the well-

known Hough transform, but never with the

fast radial symmetry transform (Loy and

Zelinsky, 2003).

 Traffic lights candidate detection is followed

by a validation process, required to exclude

false positive results. The use of TL models,

tracking, or both is the most common solution.

 Apart from (Robotics Centre of Mines

ParisTech, 2010) that provides a publicly

available annotated database of on-board

video frames taken in Paris, to the best of our

knowledge, there are no other annotated

databases for traffic lights detection.

This paper presents a TLD algorithm inspired by

the approaches followed for road sign detection, by

(Barnes and Zelinsky, 2004), (Siogkas and

Dermatas, 2006) and (Barnes et al., 2008). The Fast

Radial Symmetry (FRS) detector of (Loy and

Zelinsky, 2003) is employed in the referenced

approaches, to take advantage of the symmetrical

geometry of road signs. The symmetry and color

properties are similar in road signs and traffic lights,

so these approaches can be a good starting point.

The goal of the system is to provide a timely and

accurate detection of red and green traffic lights,

which will be robust even under adverse

illumination or weather conditions.

The proposed system is based on the CIE-

L*a*b* color space (Illumination, 1978) exploiting

the fact that the perceptually uniform a* coefficient

is a color opponency channel between red (positive

values) and green (negative values). Therefore, it is

suitable for distinction between the two prominent

classes of TLs. An image processing phase

comprising 4-connected neighborhood image flood-

filling on the positive values of a* channel is then

applied, to ensure that red traffic lights will appear

as filled circles and not as black circles with a light

background. The fast radial transform is then

utilized to detect symmetrical areas in a*. The

proposed system has been tested in various

conditions and has been qualitatively and

quantitatively assessed, producing very promising

results.

The rest of the paper is organised as follows:

Section 2 presents an overview of the proposed

system, describing in depth every module and its

functionalities. Section 3 presents experimental

results in both normal and adverse conditions, and

finally, section 4 discusses the conclusions drawn

from the experiments and suggests future work.

2 PROPOSED SYSTEM

The hardware setup for the proposed system is

similar to most of the related applications. The core

of the system is a monocular camera mounted on an

elevated position on the windshield of the moving

vehicle. The video frames shot by the camera are

processed by three cascade modules. The first one is

the pre-processing module, with a goal to produce

images in which red TLs will appear as very bright

circular blobs and green TLs will appear as very

dark circular blobs. The image obtained by the pre-

processing module is used in the traffic light

detector, which comprises a FRS transform for

various radii, followed by a detection of multiple

local maxima and minima in the top part of the

frame. The last module applies a spatiotemporal

persistency verification step to keep those candidates

that appear in multiple frames, thus minimizing false

positives.

Summing up, the proposed algorithm consists of

the following steps:

1) Frame acquisition.

2) Image pre-processing:

a) Convert RGB to L*a*b*.

b) Enhance red and green color difference.

c) Fill holes in enhanced image

3) TL candidate detection:

a) Radial symmetry detection.

b) Maxima/minima localization.

TRAFFIC LIGHTS DETECTION IN ADVERSE CONDITIONS USING COLOR, SYMMETRY AND

SPATIOTEMPORAL INFORMATION

621

4) TL candidate verification:

a) Spatiotemporal persistency check.

The steps described above are shown in Figure 1.

Figure 1: Proposed TL detection algorithm.

2.1 Pre-processing Module

Before processing each frame to detect TL

candidates, the contrast between red and green lights

and their circularity are enhanced. Ideally, red TLs

should be very bright circular blobs and green TLs

should be very dark circular blobs. In this direction,

the perceptually uniform a* channel of the CIE-

L*a*b* color space, which assigns large positive

and negative values to red and green pixels

respectively, is utilized. To further enhance the

discrimination between red and green, the pixel

values of the a* channel are multiplied by the pixel

luminosity (L channel), to produce a new channel,

called RG hereafter, defined as

(, ) (, ) (, )RGxy Lxy axy=×

, (1)

where x, y are the pixel coordinates.

This transformation results in a further increase

of the absolute value of the pixels belonging to TLs,

as they also tend to have large luminosity values. On

the other hand, green objects with very large

absolute values of a* like tree leafs, do not appear so

bright, so they are not affected on the same degree.

The same applies for most red objects, like red roofs.

The aforementioned process is demonstrated in

Figure 2. While the green TLs are still dark when

multiplying L by a* (Figure 2d), the leafs are not as

dark as in the a* channel (Figure 2c).

Ideally, the image transformation from the RGB

to the CIE-L*a*b* color space would be enough to

achieve the aforementioned goal. However, real

world video sequences containing TLs often produce

a “blooming effect”, especially in the case of red

TLs, as shown in Figure 3a. The cause of this

blooming effect can be twofold: i) red lights often

include some orange in their hue, while green TLs

(a) (b)

Figure 2: Red-green discrimination enhancement. (a) RGB

image, (b) L channel, (c) a* channel, (d) L, a*

multiplication result (RG channel).

include some blue. ii) The dynamic range of cameras

may be sensitive to very bright lights; consequently

saturated regions appear in their inner area.

To tackle this problem, we propose to combine

the red and green areas with the yellow and blue

ones, respectively. This is accomplished by

calculating the product of L and b* (YB channel):

(, ) (, ) (, )YB x y L x y b x y=×

(2)

where x, y are the pixel coordinates and then adding

the result to the RG channel to produce the RGYB

image:

()

(, ) (, ) (, )

(, ) (, ) (, ).

RGYBxy RGxy YBxy

Lxy axy bxy

=× +

(3)

The results of this process are demonstrated in

Figure 3. The “blooming effect” is shown in the red

TL on the right of the picture in Figure 3b. While it

would be expected that the value of a* to be

positive, it appears to be negative (green). The

redness is observed only in the perimeter of the TL

and could lead the FRS algorithm to detect two

symmetrical shapes with the same centre (a large

bright one and a smaller dark one). However, by

adding the YB channel (Figure 3c) to the RG result,

this problem is handled, as shown in Figure 3d.

The blooming effect, however, still remains a

problem in the case of night driving, so an additional

step is necessary. This step is a grayscale 4-

connected neighbourhood image filling (Soille,

1999) of the holes in both the bright and the dark

areas of the image. More specifically, the RGYB

image is thresholded to produce two images: one

with red and yellow pixels (RGYB > 0) and one

VISAPP 2012 - International Conference on Computer Vision Theory and Applications

622

(a) (b)

Figure 3: Handling the “blooming effect” in the morning.

(a) RGB image, (b) RG channel, (c) YB channel, (d)

RGYB channel.

with green and blue pixels (RGYB < 0). The holes in

both images are then filled, using the method

mentioned above, and then they are added to

produce the final result. This process and its results

(denoted by the red ellipses) are demonstrated in

Figure 4. The green TLs in the scene are shown as

big black circles with a concentric smaller, lighter

circle inside them (Figure 4b). This effect has been

eliminated after the aforementioned filling process,

as shown in Figure 4c.

(a) (b)

(c)

Figure 4: Handling the “blooming effect” at night. (a)

RGB image, (b) RGYB channel, (c) RGYB channel after

filling process.

2.2 Radial Symmetry Detection

The algorithm for FRS detection was first introduced

by (Loy and Zelinsky, 2002) and improved in (Loy

and Zelinsky, 2003). Its main feature is detecting

symmetrical blobs in an image, producing local

maxima in the centres of bright blobs and local

minima in the centres of dark blobs. The only

parameters that have to be defined for this process

are a radial strictness factor, a, and the radii that will

be detected.

As mentioned in section 2.1, the image

transformed by the FRS algorithm includes red and

green color opponency. This property makes it

appropriate for the FRS transform. Some examples

of the implementation of the FRS transform (for

radii from 2 to 10 pixels with a step of 2 and a=3) to

pre-processed frames of various videos are reported

in Figure 5.

Figure 5: Fast radial symmetry transforms (right column),

at day and night time (left column). Dark spots denote

green TLs, light spots denote red TLs. Local

maxima/minima detection is performed above the yellow

line (which depends on the camera placement).

The results of the FRS transform that are

demonstrated in Figure 5 show that the centres of

TLs are always within the most voted pixels of the

frame, as long as their radii are included in the

search. This means that once the FRS algorithm has

been applied, the following process is a non-maxima

suppression to detect red TLs and a non-minima

suppression to detect green TLs. However, since the

nature of the TL detection problem allows it, only

the conspicuous region of the image, i.e. the upper

part, will be used for the suppression. Since the on-

board camera was not placed identically in all the

videos used, the selection of the part of the image

that will be used for the suppression must be defined

during the system calibration. This could be

achieved automatically using horizon detection

techniques, but this is not necessary because the

areas where a TL could appear can be predicted

during the camera setup. In the examples of Figure

5, the suppression took place above the yellow line.

Up to 5 local maxima and 5 local minima are

selected, given that they lie within the range of

numbers greater than half the value of the global

maximum. Once the TL candidates are produced, the

TRAFFIC LIGHTS DETECTION IN ADVERSE CONDITIONS USING COLOR, SYMMETRY AND

SPATIOTEMPORAL INFORMATION

623

TL areas are denoted by a rectangle with coordinates

that are determined by the detected centre

coordinates, the radius and the color of the TL. More

specifically, the annotation rectangle for a green TL

starts at 6 radii up and 1.5 radii to the left of the TL

centre and has a width of 3 radii and a height of 7.5

radii. Similarly, the annotation rectangle for a red

TL starts at 1.5 radii up and 1.5 radii to the left of

the TL centre and has the same height and width as

above.

2.3 Candidate Verification

The last module of the proposed system is a TL

candidate verification process. This step is vital for

the minimization of false positives that could be

caused by numerous artefacts that resemble a TL.

Road signs, advertisements etc. can be

misinterpreted as TLs, because their shape and color

properties might look alike. However, such artefacts

usually don’t appear radially symmetrical for more

than a few frames, while the symmetry and color of

TLs are more robust to temporal scale variations.

Hence, most of them can be easily removed if the

condition of multiple appearance of a TL in

successive frames is met.

In order to satisfy spatiotemporal persistency, the

proposed method assumes that a TL will appear in

the top voted candidates for symmetrical regions in a

sequence of frames (temporal persistency), so its

centre is expected to leave a track of pixels not far

from each other (spatial persistency). Such a result is

shown in Figure 6, where the trails left by the

centres of two green TLs over a period of 30 frames

are denoted by green dots. The sparse dots on the

left belong to the other detected symmetrical objects

and do not fulfil the persistency criterion.

Figure 6: Trails left by the centres of two green TLs.

A persistent high FRS value in a small area for at

least 3 out of 4 consecutive frames is the post-

processing criterion for characterizing a pixel as

being the centre of a TL. An example of the

aforementioned process is demonstrated in Figure 7,

where the first column shows frames 1, and 2 and

the second column contains frames 3, and 4. The red

rectangles denote a red TL candidate and the yellow

ones denote a green TL candidate.

Figure 7: Four consecutive frames, TL candidates

annotated by rectangles. Non persistent candidates are

dismissed. Candidates persistent for 3 out of 4 consecutive

frames get verified.

3 EXPERIMENTAL RESULTS

The assessment of the effectiveness of the proposed

system is twofold.

First, a quantitative assessment is made, using

the publicly available, manually annotated video

sequence of (Robotics Centre of Mines ParisTech,

2010). The video sequence comprises 11179 frames

(8min 49sec) of 8-bit RGB color video, with a size

of 640x480 pixels. The video is filmed in a C3

vehicle, travelling at speeds less than 50km/h, and

the camera sensor is a Marling F-046C, with a frame

rate of 25fps and a 12mm lens, mounted behind the

rear-view mirror of the car.

Secondly, extensive qualitative testing is carried

out, using video sequences of different resolutions

shot with different cameras, under adverse

conditions in day and night time, in various

countries, downloaded off the internet (mainly from

the YouTube site), or 720x576 pixels videos at 25fps

filmed in Greek roads, using a Panasonic NV-GS180

video camera. The purpose of this experimental

procedure is to examine the system’s robustness in

changing not only the driving, weather and

illumination conditions, but also the quality of the

VISAPP 2012 - International Conference on Computer Vision Theory and Applications

624

videos and the camera setup.

The proposed system was implemented in

Matlab and was tested on a computer with a Core 2

Quad CPU at 2.83GHz, and 4GB of memory. The

code did not use any parallelization. The processing

times achieved were directly affected by the

resolution of the videos and they fluctuated from

0.1ms to 0.5ms per frame.

As far as the parameters used for the proposed

algorithm are concerned, the radii for the FRS in the

experiments were 2,4,6,8, and 10 pixels. The radial

strictness was 3, and for the spatiotemporal

persistency to hold, a candidate must appear in at

least three out of four consecutive frames, in an area

of a radius of 20 pixels.

3.1 Quantitative Results

The quantitative results of the TL detector are

estimated following the instructions given in

(Robotics Centre of Mines ParisTech, 2010), for

their publicly available, manually annotated video

sequence. These instructions stated that from the

11179 frames included in the video sequence, only

9168 contain at least one TL. However, not all the

TLs appearing in these frames are used: yellow

lights (58 instances) are excluded, as well as many

lights which are ambiguous due to heavy motion

(449 instances). Apart from these, 745 TLs that are

not entirely visible, i.e. are partially outside the

frame, are also excluded. Eliminating all these TL

instances, a total of 7916 visible and unambiguous

red or green TLs remain. The total distinct TLs that

comprise these 7916 instances are 32.

The results scored by the proposed system are

shown calculated following the same rule as (de

Charette and Nashashibi, 2009a), which states that if

a TL is detected and recognized once in the series of

frames where it appears, then it is classified as a true

positive. Hence, a false negative result is a TL never

detected in its timeline. Using these definitions, the

proposed algorithm scores a detection rate (recall) of

93.75%, detecting 30 of the total 32 TLs. The 2

missed traffic lights were green and appeared for 49

and 167 frames in the scene. In both cases, our

system detected the other green lights appearing in

the same scene, so the goal of informing the driver

can be achieved. Some detection examples from the

aforementioned database are given in Figure 8. The

total number of false positive detections is 1481, of

which 1167 are false red TL positives and 314 false

green TL detections. The red false positive instances

concern 12 different objects mistaken as red TLs,

while the number of different objects mistaken as

green TLs is 7. This means that the precision of our

system is 61.22%. This false positive rate could be

significantly reduced by using morphological cues,

like the templates utilized by (de Charette and

Nashashibi, 2009a) and (de Charette and Nashashibi,

2009b), who report precision and recall rates of up

to 98% and 97% using temporal matching. A more

direct comparison between our system and the two

aforementioned approaches is not feasible though,

due to the lack of details in the evaluation process.

Figure 8: TL detection results in daytime driving. A false

positive red TL example is show in in the bottom left

image and a false negative is shown in the bottom right.

3.2 Qualitative Results

The experimental results reported in section 3.1

show that the system appears effective and robust in

the case of urban driving in good weather and

illumination conditions. The problem of false

positives is not so persistent and could be further

improved if, as already mentioned, a TL model is

used for a final verification, or if a color constancy

module is introduced. However, the great challenge

of such systems resides in more demanding driving

conditions, including driving under rainy conditions

and driving at night time.

For this reason the proposed system is also tested

in such conditions, so that its robustness and

resilience can be examined. Various video sequences

shot from on-board cameras all around the world

have been gathered from the Internet. The ultimate

goal is to construct a database of challenging

conditions driving videos, which will be annotated

and freely distributed in the future.

3.2.1 Driving Under Rainy Conditions

The first challenging case in ADAS is driving in

rainy conditions. The difficulty present is that

TRAFFIC LIGHTS DETECTION IN ADVERSE CONDITIONS USING COLOR, SYMMETRY AND

SPATIOTEMPORAL INFORMATION

625

raindrops on the windshield often distort the driving

scene and could cause the various objects to appear

disfigured. Another problem in rainy conditions is

the partial obstruction by the windshield wiper in

various frames. For these reasons, every vision

based system should be tested under rainy

conditions, as the results produced may vary a lot

from the ones achieved in normal weather. Some

examples of successful TL detections in rainy

conditions are shown in Figure 9.

Figure 9: Successful TL detection results in rainy

conditions.

The presence of a TL for the predefined number

of frames is annotated by the inclusion of a manually

drawn TL template next to the detected TL. Some

false positive examples are annotated, but have not

appeared in previous frames, i.e. the green rectangle

in the last image on the left column of Figure 9. The

false negatives in the same picture (green TL on the

top) and in the first picture of the right column of

Figure 9 (green TL on the right) have not reached 5

consecutive detections yet and are recognised in a

later frame.

3.2.2 Driving at Night

The second important category of adverse driving

conditions is night driving. The difficulty of the

situation relies largely on the environment and the

mean luminosity of the scene. If the environment

does not include excessive noise like for example

dense advertisements or other lighting sources, the

proposed system performs well, even in urban

driving situations. Successful detections in night

driving scenarios are presented in Figure 10. Most

TLs are successfully detected, even when their glow

makes it very difficult to distinguish shape and

morphology cues.

Figure 10: TL detection results in urban night driving.

3.2.3 Known Limitations of the TLD

A very common question when dealing with

computer vision applications is whether there are

problems that they cannot solve. The proposed

method is by no means flawless and can produce

persistent errors in some situations. The main

problems that can be pinpointed are the following:

i) The system produces some false positive results

that cannot be easily excluded, unless it is used in

correlation to other computer vision modules like a

vehicle detector or a road detector. An example of

such false positives is illuminated vehicle tail lights,

or turn lights, as in Figure 11a. This image also

includes a false positive result that is excluded in

next frames. ii) The proposed system fails

completely in cities like New York (Figure 11b),

where the visual information is extremely dense and

the TLs are lost in the background.

(a) (b)

Figure 11: Examples of temporary or permanent failures

of the proposed system.

4 CONCLUSIONS

In this paper, we have proposed a novel automatic

algorithm for traffic lights detection using a

monocular on-board camera. The algorithm uses

VISAPP 2012 - International Conference on Computer Vision Theory and Applications

626

color, symmetry and spatiotemporal information to

detect red and green traffic lights in a fashion

resilient to weather, illumination, camera setup and

time of day. It utilizes a CIE-L*a*b* based color

space with a holes filling process to enhance the

seperability of red and green traffic lights. A fast

radial symmetry transform is then used to detect the

most symmetrical red and green regions of the upper

part of the image, producing the TL candidates.

Finally, a spatiotemporal persistency criterion is

applied, to exclude many false positive results. The

algorithm has been experimentally assessed in many

different scenarios and conditions, producing very

high detection rates, even in very adverse conditions.

Future work will be directed towards embedding

a tracking module to the algorithm to minimize the

false negative results and a color consistency module

to further reduce false positives. Furthermore, the

combination of the TL detector with other ADAS

modules like vehicle, sign and road detection will be

explored, so that a complete solution for driver

assistance is proposed.

REFERENCES

Barnes, N., & Zelinsky, A. (2004). Real-time radial

symmetry for speed sign detection (pp. 566–571).

Presented at the 2004 IEEE Intelligent Vehicles

Symposium, IEEE.

doi: 10.1109/IVS.2004.1336446

Barnes, N., Zelinsky, A., & Fletcher, L. S. (2008). Real-

time speed sign detection using the radial symmetry

detector. Intelligent Transportation Systems, IEEE

Transactions on, 9(2), 322–332.

Chunhe Yu, Chuan Huang, & Yao Lang. (2010). Traffic

light detection during day and night conditions by a

camera (pp. 821-824). Presented at the 2010 IEEE

10th International Conference on Signal Processing

(ICSP), IEEE. doi:10.1109/ICOSP.2010.5655934

de Charette, R., & Nashashibi, F. (2009a). Real time

visual traffic lights recognition based on Spot Light

Detection and adaptive traffic lights templates (pp.

358-363). Presented at the 2009 IEEE Intelligent

Vehicles Symposium, IEEE. doi:10.1109/IVS.

2009.5164304

de Charette, R., & Nashashibi, F. (2009b). Traffic light

recognition using image processing compared to

learning processes (pp. 333-338). Presented at the

IEEE/RSJ International Conference on Intelligent

Robots and Systems, 2009. IROS 2009, IEEE.

doi:10.1109/IROS.2009.5353941

Illumination, I. C. on. (1978). Recommendations on

uniform color spaces, color-difference equations,

psychometric color terms. Bureau central de la CIE.

In-Hak, J., Seong-Ik, C., & Tae-Hyun, H. (2006).

Detection of Traffic Lights for Vision-Based Car.

Traffic, 4319/2006, 682-691. doi:10.1007/

11949534_68

Lindner, F., Kressel, U., & Kaelberer, S. (2004). Robust

recognition of traffic signals (pp. 49- 53). Presented at

the 2004 IEEE Intelligent Vehicles Symposium, IEEE.

doi:10.1109/IVS.2004.1336354

Loy, G., & Zelinsky, A. (2002). A fast radial symmetry

transform for detecting points of interest. Computer

Vision—ECCV 2002, 358–368.

Loy, G., & Zelinsky, A. (2003). Fast radial symmetry for

detecting points of interest. IEEE Transactions on

Pattern Analysis and Machine Intelligence, 959–973.

Omachi, M., & Omachi, S. (2009). Traffic light detection

with color and edge information (pp. 284-287).

Presented at the 2nd IEEE International Conference on

Computer Science and Information Technology, 2009.

ICCSIT 2009, IEEE. doi:10.1109/ICCSIT.2009.

5234518

Robotics Centre of Mines ParisTech. (2010, May 1).

Traffic Lights Recognition (TLR) public benchmarks

[La Route Automatisée]. Retrieved October 6, 2011,

from http://www.lara.prd.fr/benchmarks/trafficlights

recognition

Siogkas, G. K., & Dermatas, E. S. (2006). Detection,

Tracking and Classification of Road Signs in Adverse

Conditions (pp. 537-540). Presented at the IEEE

Mediterranean Electrotechnical Conference,

MELECON 2006, Malaga, Spain.

doi:10.1109/MELCON.2006.1653157

Soille, P. (1999). Morphological image analysis :

principles and applications with 12 tables. Berlin

[u.a.]: Springer.

Yehu Shen, Ozguner, U., Redmill, K., & Jilin Liu. (2009).

A robust video based traffic light detection algorithm

for intelligent vehicles (pp. 521-526). Presented at the

2009 IEEE Intelligent Vehicles Symposium, IEEE.

doi:10.1109/IVS.2009.5164332

TRAFFIC LIGHTS DETECTION IN ADVERSE CONDITIONS USING COLOR, SYMMETRY AND

SPATIOTEMPORAL INFORMATION

627