EVALUATION OF A ROAD SIGN PRE-DETECTION SYSTEM BY

IMAGE ANALYSIS

Philippe Foucher, Pierre Charbonnier and Houssem Kebbous

ERA 27 LCPC, Laboratoire des Ponts et Chauss

ees, 11 rue Jean Mentelin, B.P. 9, 67035 Strasbourg, France

Keywords:

Image analysis, Road sign detection, Road sign inventory, Color, Shape, Symmetry, Evaluation.

Abstract:

In this paper, we introduce a pre-detection algorithm dedicated to French danger-warning and prohibitory

road signs. The proposed method combines color, shape, location and symmetry features to select among

large image databases, a small subset of pictures that probably contain road signs. We report the results of

a systematic experimental assessment that we performed on ﬁve image databases, comprised of more than

26,000 images, covering 176 km and containing 371 trafﬁc signs, among which a non-negligible amount

(about 5% in average) is damaged. The experiments show that about 10% images of the sequences are selected

and more than 87% trafﬁc signs are detected. The missed objects always correspond to dirty, worn-out or badly

oriented signs that would be difﬁcult to detect even for a human operator.

1 INTRODUCTION

In this paper, we address the problem of road sign

inventory. In our application, evenly spaced still-

images of the roadway and its close surroundings (see

e.g. Fig. 1) are collected by inspection vehicles along

itineraries, typically one image every 5 or 10 meters.

These image databases are then processed off-line by

human operators, which is fastidious and prone to

oversights. Our research aims at proposing image

analysis tools for automating the operator’s task. In

particular, it may be useful to select, among all the

data set, the small number (typically, 5 to 10 %) of im-

ages that probably contain road signs. We propose in

this paper such a pre-detection procedure, that relies

on color, shape, location and symmetry features. The

algorithm is described in the case of French danger-

warning and prohibitory signs, which are red, but may

readily be extended to other types of colored road

signs. The second contribution of the paper is that

we perform a systematic and objective evaluation of

the pre-selection procedure on large image databases,

taken by various cameras.

The paper is organized as follows. We ﬁrst pro-

pose a brief review of related work (Sect. 2). Then,

Sect. 3 describes the three steps of the pre-detection

algorithm. Sect. 4 presents the experimental setup

and in Sect. 5 we comment the experimental results.

2 RELATED WORK

Trafﬁc sign detection and recognition by image anal-

ysis are useful for many applications such as road sign

inventory, driving assistance or autonomous systems.

In the last two decades, many research teams have

been interested in this topic, see e.g. (Barnes et al.,

2008) for a recent survey. Detection methods gener-

ally rely on the fact that road signs are manufactured,

standardized objects, whose shape, dimensions, color

and location are ﬁxed by norms. However, detection

systems encounter typical problems of computer vi-

sion in uncontrolled environments, such as: variations

in lighting conditions (brightness, reﬂections, shad-

ows...), geometric distortions (rotations), complex-

ity of road scenes (neighboring objects with similar

color, occlusions), object’s wear and tear. In many re-

searches, the detection system uses a combination of

color and shape information. Color-based segmenta-

tion is commonly applied ﬁrst, to extract pixels corre-

sponding to the color of the road sign. The Hue com-

ponent, which is not sensitive to lighting changes, is

usually used for this task (Shaposhnikov et al., 2002;

de la Escalera et al., 2003). However, an efﬁcient

alternative is to consider the normalized component

R/(R + G + B) which is also independent on light-

ing changes (Dutilleux and Charbonnier, 2007). Sec-

ondly, the connected components are selected accord-

ing to some shape parameters, e.g. size of the ob-

ject area (Maldonado-Bascon et al., 2007), geometri-

362

Foucher P., Charbonnier P. and Kebbous H. (2009).

EVALUATION OF A ROAD SIGN PRE-DETECTION SYSTEM BY IMAGE ANALYSIS.

In Proceedings of the Fourth International Conference on Computer Vision Theory and Applications, pages 362-367

DOI: 10.5220/0001801003620367

 SciTePress

cal properties of edge segments (Piccioli et al., 1996)

(note that in (Piccioli et al., 1996), the robustness of

the shape detector to occlusions is not discussed).

Some authors directly work with grey-level im-

ages to avoid the difﬁculties of color-based segmenta-

tion. For instance, (Loy and Barnes, 2004) adapt the

Fast Radial Symmetry detector described by (Loy and

Zelinsky, 2003) to the detection of regular polygon.

This paper presents a systematic evaluation of the

performance of the pre-detection algorithm on differ-

ent sets of sequences. In the literature, the evaluation

of road sign detectors often concerns only a limited

number of images and/or situations. To our knowl-

edge, quantitative evaluation on a large database was

only performed in (Priese et al., 1993). In this project,

a detection and recognition algorithm was tested on

20000 images and the performance were given in

terms of true detections and false alarms. More re-

cently, (Barnes et al., 2008) uses the Receiving Op-

erating Characteristic (ROC) curves to perform a sys-

tematic evaluation of a detection and recognition al-

gorithm.

3 METHODOLOGY

The symmetry algorithms applied to large-size color

image implies a prohibitive computation time. Cur-

rently, computing in Matlab the polygonal symmetry

transform on a single 1280 × 1024 image takes 1273

seconds. Even if an implementation in C may acceler-

ate the process, a pre-selection of connected compo-

nents based on color and shape information seems to

be appropriate. Therefore, our complete pre-detection

algorithm consists of three steps: (1) color-based seg-

mentation (2) selection of resulting connected com-

ponents according to the size, shape and position of

the candidates (3) selection by a symmetry criterion.

3.1 Color

In most countries, warning and prohibitory trafﬁc

signs have a thick red border, which is a very discrimi-

nant characteristic. In the ﬁrst step of the method, pix-

els are (individually) classiﬁed into red/non-red. An

(R,G,B) pixel is considered as red if (Dutilleux and

Charbonnier, 2007):

R > α(G + B)

R − max(G,B) > β[max(G,B) − min(G, B)]

(1)

The ﬁrst expression selects pixels whose normalized

red component R/(R + G + B) dominates. Normal-

ization provides robustness to lighting changes. The

second expression ﬁlters out pixels that tend either

towards yellow or towards magenta (Dutilleux and

Charbonnier, 2007). The threshold β has not been em-

pirically determined but we found the rule β = 2α sat-

isfactory for all the sequences we encountered. More-

over, our experience is that the best values for α, in

terms of classiﬁcation performance, range between

0.55 and 0.6, the former being a less selective value

than the latter. An exemple of red classiﬁcation map

is shown on ﬁg. 1.

3.2 Size, Shape and Position

After pixel classiﬁcation, red connected components

are examined according to several geometric parame-

ters, which reduces the number of candidates.

1. Connected components are ﬁrst ﬁltered according

to their area. The thresholds of course depend on

the characteristics of the image acquisition sys-

tem.

2. The position of the center of mass of the con-

nected component must be in a region-of-interest

which correspond to the standard road sign setting

in right-hand trafﬁc countries (see Fig. 1).

3. The eccentricity of the components must be

smaller than a threshold, whose value results from

a trade-off between sensitivity to elongated ob-

jects (e.g. poles), that may result in false alarms,

and sensitivity to partial occlusions, that decrease

the capacity of detection. This value is empiri-

cally determined, as explained in Sec. 5.

4. Red connected components that correspond to

road signs are relatively empty objects. The ex-

tent of a connected component corresponds to the

ratio between its area and the area of its bound-

ing box. Bounds on the extent are set by taking

margins around the values deduced from norms.

This allows distinguishing yield signs from dan-

ger signs, for example.

5. In the case of warning signs, the orientation of

edges is a discriminant characteristic. The angles

between the lateral sides of the connected compo-

nent and the horizontal are checked. The valid-

ity range accounts for a certain margin around the

theoretical value of 60

◦

, to provide some ﬂexibil-

ity with respect to badly oriented trafﬁc signs.

3.3 Symmetry

The last selection step checks if the remaining con-

nected components are symmetric. To this end, we

compute the Fast Radial Symmetry transform (or its

extension to polygonal objects) of the luminance im-

age, on a window that corresponds to the bounding

EVALUATION OF A ROAD SIGN PRE-DETECTION SYSTEM BY IMAGE ANALYSIS

363

Figure 1: Road scene example (left); red/non-red pixel clas-

siﬁcation (right). The dotted polygon shows the search re-

gion for the selection of connected components. Pixels clas-

siﬁed as red are shown in black.

box of each selected component. Checking the sym-

metry of an object simply amounts to thresholding

the resulting vote image. When all connected com-

ponents have been examined, the image is selected or

not, according to the number of remaining objects.

4 EXPERIMENTAL SETUP

4.1 Evaluation Data Sets

In this experimental validation, we consider real-

world image sequences acquired by frontal cameras

mounted on top of inspection vehicles. The charac-

teristics of the test sequences are given in Tab. 1. Im-

ages are taken every 5 or 10 meters, with several kinds

of CCD cameras under various, uncontrolled lighting

conditions. Note that the evaluation databases com-

prises about 26,000 images and covers more than 175

km of rural roads, urban roads and highways.

Table 1: Characteristics of the test sequences.

Seq. Number

of images

Step Camera Size (pixels) Size

(Ko)

S1 3436 5 m Basler A101 1280 ×1024 156

S2 3436 5 m Basler A101 1280 ×1024 156

S3 10420 5 m Marlin F145-C2 1280 × 1024 95.6

S4 890 10 m Pike F210-C 1920 × 1080 209

S5 8103 10 m Pike F210-C 1920 × 1080 223

4.2 Ground Truth

The test sequences were manually processed using

a specially designed software to establish a ground

truth, i.e. to list those images where road signs are

visible. Of course, a particular instance of road sign is

most of the time visible in several images. We there-

fore distinguish between the number of trafﬁc signs

and the number of images that show trafﬁc signs.

These numbers are given in Tab. 2 (WS stands for

warning signs and PS stands for prohibitory signs).

Table 2: Ground truth (manual analysis of test sequences).

Seq. Number

of WS

Numb. of im-

ages with WS

Number

of PS

Numb. of im-

ages with PS

S1 25 123 18 93

S2 12 81 5 44

S3 64 383 61 445

S4 9 36 18 63

S5 70 377 89 540

Total 180 1000 191 1185

We note that among the total number of road

signs, some of them (about 5% in average) are either

worn-out, dirty, badly oriented, or non-standard.

4.3 Evaluation Metrics

To evaluate the performances of the pre-detection sys-

tem, we use ROC curves, which plot the True Positive

Rate (TPR) vs the False Positive Rate (FPR). Each

point on a ROC curve corresponds to a particular set-

ting of the detector’s parameters.The False Positive

Rate is given by:

FPR =

(2)

where FP is the number of images that contain at least

one false alarm. Note that false alarms may occur in

images that show a road sign, as soon as the detected

component does not ﬁt in the ground-truth bounding

box. Hence, N is the total number of images in the

sequence.

In our application, it sufﬁces to detect a sign once

over the series of images it appears in. Therefore, we

deﬁne two different true positive rates: the ﬁrst one,

TPR

, for evaluating the detection of images with traf-

ﬁc signs and the second one, TPR

, for assessing the

detection of road signs. These are deﬁned as:

T PR

T P

and T PR

T P

(3)

where TP

is the number of selected images with traf-

ﬁc signs, P

is the actual number of images with trafﬁc

signs, TP

is the number of detected trafﬁc signs and

is the actual number of trafﬁc signs.

5 EXPERIMENTAL RESULTS

In this section, the results on our ﬁve test sequences

are presented and discussed. Some parameters are a

priori ﬁxed from theoretical characteristics of trafﬁc

signs. For example, the minimum size of connected

component is 200 pixels (corresponding to a 30 pixels

high object) and the maximum size is 2.5% the image

size (the size of the biggest observed sign corresponds

VISAPP 2009 - International Conference on Computer Vision Theory and Applications

364

to 1.5% the image size). The upper extent threshold

is ﬁxed at 0.35 for the warning signs. The prohibitory

signs category includes the “no entry” sign, so the ex-

tent threshold is ﬁxed at 0.65 at the risk of a higher

false alarm rate. The lower threshold is not used, to

accept occluded objects. The validity range for angle

in the case of warning signs is set to [50

◦

,70

◦

Sequences S1 and S4 are used to set the value α

and the thresholds of eccentricity and symmetry. The

algorithm is then tested on sequences S2, S3 and S5.

5.1 Considering Color and Eccentricity

In this series of experiments, the symmetry selection

is disabled, to focus on the effect of colour and ec-

centricity criteria. The inﬂuence of the red classi-

ﬁcation parameter, α and of the eccentricity thresh-

old is illustrated by the ROC curves plotted on ﬁg. 3

for prohibitory signs and on ﬁg. 5 for warning signs.

In these experiments, the eccentricity threshold varies

over the range [0.6, 1], and α is ﬁxed to 0.55 or 0.6.

For both categories, α = 0.55 naturally yields more

false alarms than α = 0.6 but it may also be noticed

that the number of true positive is lower. In fact, since

more pixels are classiﬁed as red when α = 0.55, so

the red component that corresponds to the sign border

may connect with background elements and then, be

ﬁltered off by the shape analysis stage. The inﬂuence

of the color parameter might be worth a more thor-

ough study. However, the value of α = 0.6 will be

considered in the remaining of the paper.

Concerning eccentricity, the best detection scores

are obtained for a threshold of 1, but at the price of

a high false alarm rate. When the test becomes more

selective, typically for thresholds under 0.85, partly

occulted objects are more difﬁcultly detected. No-

tice that, for prohibitory signs, varying the eccentric-

ity threshold from 0.85 to 0.9 improves the true de-

tection rate faster than the false alarm rate. Moreover,

true detection rate almost reaches its maximum for a

threshold of 0.9. Hence we retain this value for the

rest of our experiments. Similarly, we chose the value

of 0.85 in the case of warning signs.

The ROC curves for S4 are easily distinguishable

from those corresponding to S1. The false alarm rate

is much higher in S4 than in S1. Possible explanations

are different cameras, colorization settings, compres-

sion rates and nature of scenes (S4 being more urban

than S1, for example). This shall be investigated in a

near future.

It may be noticed that for prohibitory signs, a

100% true positive rate (TPR

) is never reached.

Some non-detected trafﬁc signs correspond to far-

away signs that may be detected when they appear

closer. When TPR

is considered, the detection rate

reaches 94.4% for the S1 sequence (1 road sign is

never detected) and 83.3% for the S4 sequence (3 road

signs are missed). Non-detected signs correspond to

a severely worn-out sign (see Fig. 2(e)) and to signs

parallel to the axis of the road. In the case of warning

signs, non-detected objects in S1 correspond to old-

fashioned, non-standard, worn-out or temporary signs

and to one sign with yellow ﬂashes (see Fig. 2(f)).

Note that all warning signs are detected in S4.

5.2 Inﬂuence of Symmetry

In this experiment, the symmetry threshold varies

over the range [0, 70]. The value 0 corresponds to

no symmetry selection. The resulting ROC curves

are plotted on ﬁg. 4 and ﬁg. 6. Examples of good

detection, including difﬁcult scenes, are presented on

ﬁg. 2(a)-(d). Selection based on symmetry has a no-

ticeable impact on the false alarm rate. This effect

is less obvious for warning signs than for prohibitory

signs due to a higher extent threshold. However, the

initial false positive rate was already weak for warn-

ing signs, thanks to a more selective geometrical cri-

terion, namely selection according to angles. Re-

call that the extent criterion is also more efﬁcient for

warning signs than for prohibitory ones. This raises

the question of the utility of selection on symmetry,

which is rather costly in terms of computation time

for polygonal objects, since the number of false posi-

tive may already be low for certain sequences.

In general, selection on symmetry tends to discard

the ﬁrst image of the series a road signs appears in,

but this has no impact in our application. For circular

signs, 3 worn-out “no parking” signs are lost in S1. In

France, these signs have a red ring and a blue interior.

The lack of contrast between the two colors explain

the non-detection. This might be overcome by using

other modalities than luminance. For warning signs,

selection on symmetry does not decrease the true pos-

itive rate TPR

, except for a high threshold.

In conclusion, selection on symmetry allows

strongly decreasing the number of false positives for

radial objects. Its interest appears more clearly than

for polygonal road signs.

5.3 Validation

The above experiments have lead to the choice of 0.9

for the eccentricity and 40 (resp. 30) for the symmetry

threshold in the case of prohibitory (resp. warning)

signs. We also take α = 0.6, and the other parameters

are set as explained at the beginning of sec. 5. In this

validation experiment,we apply the detector with the

EVALUATION OF A ROAD SIGN PRE-DETECTION SYSTEM BY IMAGE ANALYSIS

365

same set of parameters to sequences S2, S3 and S5.

The results are gathered in tab. 3 for prohibitory road

signs and in tab. 4 for warning signs.

Table 3: Detection of prohibitory signs (PS).

Seq. TPR

TPR

FPR Rate of selected images

S1 0.67 0.78 0.05 0.07

S4 0.77 0.83 0.175 0.23

S2 1 1 0.015 0.03

S3 0.84 0.9 0.08 0.12

S5 0.86 0.95 0.07 0.12

Table 4: Detection of warning signs (WS).

Seq. TPR

TPR

FPR Rate of selected images

S1 0.77 0.84 0.02 0.05

S4 0.61 1 0.185 0.21

S2 0.88 1 0.01 0.03

S3 0.63 0.875 0.03 0.05

S5 0.67 0.95 0.07 0.11

The results on S2 are similar to those obtained on

S1, which was expected since both are part of a single

sequence. The results on S5 are even better than those

on S4. In fact, the same camera was used, but the ac-

quisition conditions were different and the itinerary is

less urban. Results on S3, i.e. with a different cam-

era, that was not involved in the parameter selection

process, are of the same quality, which is very satis-

factory. The true positive rate is lower then 90% on

S3 for warning signs, but this sequence contains more

damaged signs and compression artifacts than others.

6 CONCLUSIONS

In this paper we proposed a method for pre-selecting

images that potentially contain road signs in large im-

age databases. The method quickly ﬁlters objects ac-

cording to colour, shape, location and symmetry cri-

teria. We applied this algorithm on ﬁve large real-

world image sequences featuring damaged and badly

oriented trafﬁc signs. In general, about 10% images

are selected, which fulﬁlls the requirements of our ap-

plication, and would also be a good starting point for

more automatic detection and recognition algorithms

(Dutilleux and Charbonnier, 2007) that will hopefully

further reduce the false alarm rate. The true detection

rates (TPR

) are quite good on this kind of difﬁcult

image sequences: the missed objects always corre-

spond to dirty, worn out or badly oriented trafﬁc signs

that sometimes would not be detected by human op-

erators either. We notice that the symmetry algorithm

(Barnes et al., 2008) directly applied on the whole im-

age may not detect the damaged trafﬁc signs. More-

over, our complete algorithm takes 3500 sec. to pro-

cess the whole S1 sequence (3436 images) while the

symmetry method takes 1200 sec. for one image.

REFERENCES

Barnes, N., Zelinsky, A., and Fletcher, L. (2008). Real-time

speed sign detection using the radial symmetry detec-

tor. IEEE Transactions on Intelligent Transportation

Systems, 9(2):322–332.

de la Escalera, A., Armingol, J. M., and Mata, M. (2003).

Trafﬁc sign recognition and analysis for intelligent ve-

hicles. Image and Vision Computing, 21(3):247–258.

Dutilleux, G. and Charbonnier, P. (2007). M

etaheuristiques

biologiques pour la d

etection de la signalisation

routi

ere. In Siarry, P., editor, Optimisation en traite-

ment du signal et de l’image, Trait

e IC2, s

erie traite-

ment du signal et de l’image, chapter 10, pages 271–

294. Hermes.

Loy, G. and Barnes, N. (2004). Fast shape-based road sign

detection for a driver assistance system. In IEEE/RSJ

International Conference on Intelligent Robots and

Systems (IROS’04), volume 1, pages 70–75, Sendai,

Japan.

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

detecting points of interest. IEEE Transactions on Pat-

tern Analysis and Machine Intelligence, 25(8):959–

973.

Maldonado-Bascon, S., Lafuente-Arroyo, S., Gil-Jimenez,

P., Gomez-Moreno, H., and F.Lopez-Ferreras (2007).

Road-sign detection and recognition based on support

vector machines. IEEE Transactions on Intelligent

Transportation Systems, 8(2):264–278.

Piccioli, G., Micheli, E. D., Parodi, P., and Campani, M.

(1996). A robust method for road sign detection and

recognition. Image and Vision Computing Journal,

14:209–223.

Priese, L., Rehrmann, V., Schian, R., and Lakmann, R.

(1993). Trafﬁc sign recognition based on color image

evaluation. In IEEE Intelligent Vehicles Symposium,

pages 95–100, Tokyo, Japan.

Shaposhnikov, D., Lubov, N., Podladchikova, L., Golovan,

A., Shevtsova, N., Hong, K., and Gao, X. (2002).

Road sign recognition by single positioning of space-

variant sensor window. In 15th International Con-

ference on Vision Interface, pages 213–217, Calgary,

Canada.

VISAPP 2009 - International Conference on Computer Vision Theory and Applications

366

(a) (b) (c) (d) (e) (f)

Figure 2: Detection of trafﬁc signs : (a) simple case; (b)old WS ; (c) dirty sign border; (d) occulted WS; (e) severely worn-out

PS ; (f) non detected WS with ﬂashes. Detected objects appear in red, superimposed on the original image.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.3

0.4

0.5

0.6

0.7

0.8

0.9

False Positive Rate

True Positive Rate

0.6

0.65

0.9

0.95

0.6

0.65

0.75

0.8

0.85

0.9

0.95

0.6

0.65

0.7

0.75

0.8

0.85

0.9

0.95

0.6

0.65

0.7

0.75

0.8

0.85

0.9

S1 sequence, color 0.6

S1 sequence, color 0.55

S4 sequence, color 0.6

S4 sequence, color 0.55

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.3

0.4

0.5

0.6

0.7

0.8

0.9

False Positive Rate

True Positive Rate

0.6

0.7

0.75

0.65

0.7

0.8

0.9

0.95

0.6

0.65

0.7

0.75

0.8 0.9

0.6

0.65 0.7

0.8

0.85

0.95

S1 sequence, color 0.6

S1 sequence, color 0.55

S4 sequence, color 0.6

S4 sequence, color 0.55

Figure 3: Inﬂuence of color and eccentricity parameters on

the detection of PS, using TPR

(top) and TPR

(bottom);

the eccentricity threshold varies over [0.6,1].

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.3

0.4

0.5

0.6

0.7

0.8

0.9

False Positive Rate

True Positive Rate

010

3040

010

3040

S1 sequence (TPR)i , color 0.6 , eccentricity 0.9

S1 sequence (TPR)rs, color 0.6 , eccentricity 0.9

S4 sequence (TPR)i , color 0.6 , eccentricity 0.9

S4 sequence (TPR)rs , color 0.6 , eccentricity 0.9

Figure 4: Inﬂuence of symmetry on the detection of PS; the

symmetry threshold varies over [0,70].

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.3

0.4

0.5

0.6

0.7

0.8

0.9

False Positive Rate

True Positive Rate

0.6

0.95

0.6

0.65

0.75

0.8

0.85 0.9

0.95

0.6 0.7

0.8

0.85

0.9

0.95

S1 sequence, color 0.6

S1 sequence, color 0.55

S4 sequence, color 0.6

S4 sequence, color 0.55

0.85

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.3

0.4

0.5

0.6

0.7

0.8

0.9

False Positive Rate

True Positive Rate

0.6

0.95

0.6

0.75

0.6 0.85

0.75

S1 sequence, color 0.6

S1 sequence, color 0.55

S4 sequence, color 0.6

S4 sequence, color 0.55

Figure 5: Inﬂuence of color and eccentricity parameters on

the detection of WS, using TPR

(top) and TPR

(bottom);

the eccentricity threshold varies over [0.6,1].

0 0.05 0.1 0.15 0.2 0.25 0.3

0.3

0.4

0.5

0.6

0.7

0.8

0.9

False Positive Rate

True Positive Rate

S1 sequence (TPR)i, color 0.6, eccentricity 0.85

S1 sequence (TPR)rs, color 0.6, eccentricity 0.85

S4 sequence (TPR)i, color 0.6, eccentricity 0.85

S4 sequence (TPR)rs, color 0.6, eccentricity 0.85

Figure 6: Inﬂuence of symmetry on the detection of WS; the

symmetry threshold varies over [0,70]. Notice the reduced

abscissa range.

EVALUATION OF A ROAD SIGN PRE-DETECTION SYSTEM BY IMAGE ANALYSIS

367