Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

2014 27th SIBGRAPI Conference on Graphics, Patterns and Images

Helmet Detection on Motorcyclists Using Image


Descriptors and Classiers
Romuere Silva Kelson Aires and Rodrigo Veras
Departamento de Sistemas de Informaca o Departamento de Computaca o
Universidade Federal do Piau - UFPI Universidade Federal do Piau - UFPI
Picos, Brasil Teresina, Brasil
Email: romuere@ufpi.edu.br Email: {kelson, rveras}@ufpi.edu.br

AbstractMotorcycle accidents have been rapidly growing 66% of motorcyclists used the helmet according to the law [5].
throughout the years in many countries. Due to various The Centers for Disease Control and Prevention in USA
social and economic factors, this type of vehicle is becoming presented a report in 2012 where the ndings indicated that
increasingly popular. The helmet is the main safety equipment
of motorcyclists, however many drivers do not use it. The main 12% of fatally injured motorcyclists were not wearing helmets
goal of helmet is to protect the drivers head in case of accident. in states with universal helmet laws1 . In states that adopt
In case of accident, if the motorcyclist does not use can be fatal. partial helmet law2 the number increased to 64%. Finally, the
This paper aims to propose a system for detection of motorcyclist number was 79% in states without a helmet law [6]. This
without helmet. For this, we have applied the circular Hough statistics show how is important the helmet use.
transform and the Histogram of Oriented Gradients descriptor
to extract the image attributes. Then, the MultiLayer Perceptron Over the past years many works were carried out in
classier was used and the obtained results were compared with trafc analysis, including vehicle detection and classication,
others algorithms. Trafc images were captured by cameras from and helmet detection. Intelligent trafc systems was usually
public roads and constitute a database of 255 images. Indeed, the implemented using vision computer algorithms, such as:
algorithm step regarding the helmet detection accomplished an
background and foreground image detection to segment
accuracy rate of 91.37%.
moving objects in scene and image descriptors to extract
Keywords-Helmet; Feature extraction; Classication; features. Computational intelligence algorithms are used too,
like machine learning algorithms to classier the objects. Next,
I. I NTRODUCTION some related works to helmet detection are presented.
Motorcycles are one of the most mode of transport used in Wen et al. [7] suggested a circle arc detection method based
the world. It is caused by the low prices and low operation upon the circular Hough transform. They applied it to detect
cost in comparison with another vehicles. In the last decade, helmet on the surveillance system of the Automatic Teller
it was observed an increase in the number of motorcycle Machine. The weakness of this work is that they only use
accidents. In 2011, according to the National Department of geometric features to verify if any safety helmet exists in
Transport Infrastructure (DNIT), Brazil had a total of 34, 635 the scene. Geometric features are not enough to nd helmet.
motorcycles involved in accidents [1]. Based on another The head can be mistaken with a helmet because both had a
research [2], the number of deaths in the same year was circular-like form.
14.666. In another paper Chiu et al. [8] proposed a computer
According to statistics of 2012 [3], Thailand had more than vision system aims to detect and segment motorcycles partly
17 millions motorcycles ofcially registered. Due to a large occluded by another vehicle. A helmet detection system is
number of vehicles, a high number of accidents is expected. used, and the helmet presence determines that there is a
According to the World Health Organization report [4] motorcycle in a scene. In order to detect the helmet presence,
published in 2013 about road safety, the trafc-related death the image edges are computed in the possible helmet region.
rate in Thailand was 38.1 per 100 thousand peoples, and 118.8 The Canny edge detector [9] is used. The quantities of edge
per 100 thousand motor vehicles. The same report said that points which are similar to a circle dene a helmet region.
Thailand has the third higher trafc-related mortality rate. In The method needs so much information (helmet radius, camera
Brazil, the average rate was 22.5, and in USA this value was angle and height) that must be provided by user.
11.4. In addition, the report said that around 25% of deaths in In another paper Chiverton [10] described and tested a
trafc-related accidents in Brazil envolved motorcyclists. system for tracking and automatic classication of riders with
The main safety equipment of motorcyclist is the helmet. and without helmet. The system uses the SVM classier,
The helmet protects the motorcyclist against accidents. which is trained with the attribute vectors of the Histogram
Although the helmet use is mandatory in many countries, there
are motorcyclists that do not use it or use incorrectly. The 1 The helmet use is mandatory for all motorcyclists.
Department of Transportation in USA said that, in 2011, only 2 Only required specic groups, usually young riders, to wear helmet.

1530-1834/14 $31.00 2014 IEEE 141


DOI 10.1109/SIBGRAPI.2014.28
of Oriented Gradients (HOG) descriptor [11]. The HOG classier to separate the objects in two classes: motorcycle
descriptor is executed with the Sobel operator [12] for and non-motorcycle. The objects are classied only into two
calculating the edges with a neighborhood of 3 3 pixels. groups just because it is enough to know that each object
The algorithm proposed in [13] is used to calculate the corresponds to a motorcycle or not.
background. Still photographs and individual image frames In the moving objects detection and classication steps we
from the video data are used on the histograms extraction. The use the same strategy of [16]. Thus, we do not show the results
method achieved a total accuracy rate of 83% for classication of this classication.
of motorcycles and 85% for helmet detection. It is noticed that
the amount of images in the testing phase is small. C. Helmet Detection
The most recent work on helmet detection was proposed by After the motorcycle detection, we made the helmet
Waranusast et al. [14]. In this work the classication is made detection step is perfomed. This step is divided in four phases:
by geometric and color information. These attributes are used 1) Region of interest (RoI) determination;
by the K-Nearest Neighbors (KNN) [15] to classify images 2) Sub-window calculation;
of people based on helmet use. Moreover, the system counts 3) Features extraction;
the number of passengers on the motorcycle. According to the 4) Image classication.
authors, the passenger counter had a total of 83.82% accuracy. The RoI aims to compute a small region to reduce the
In the helmet detection phase, the accuracy was 89%. It is computational cost and improve the helmet search. After, it
noteworthy that, for this phase, the images of the head were is obtained a sub-window that contains the motorcyclist head
manually cut. The images were captured perpendicular to the region. Next, the features were extracted from the sub-window
camera, i.e. the images are taken of the side of the motorcycle. the classiers are used.
This makes the system defective, since it is impossible to
capture the license plate of the vehicle in that position. III. R EGION OF I NTEREST D ETERMINATION
However, using these images, the identication of more than The RoI determination is an important step in our problem.
one person in the motorcycle is facilitated. If the images had It is used to search the helmet. The motorcyclist head must
been taken by another angle, probably one person on the be inside the RoI. Using RoI decreases the computational
motorcycle would override the other, generating occlusion. cost the search area. In our database, all images present the
II. S YSTEM OVERVIEW motorcyclist head region inside the RoI. The RoI was specied
at the image top and corresponds to 1/5 of the image height.
This paper presents a method to detect motorcyclists without This value was chosen empirically. Figure 1 shows the RoI
helmet on public roads. The system is divided in three determination.
steps: 1) Moving objects segmentation; 2) Moving objects
classication; 3) Helmet detection.

A. Moving Objects Segmentation
The moving objects segmentation in scene allows that only
objects of interest in the image are evaluated. This causes a
reduction in processing time due to the fact that the algorithm
works in a small area of the image. It also reduces the 
probability of false-positives in classiers.
A background (BG) image is necessary to segment the
moving objects. From BG is possible know the moving pixels.
We used a video camera to capture the trafc images. In
environments where the static objects change during the time
(parked vehicles along the roads, changes position of the
shadow, etc), the algorithms to calculate adaptive BGs are Fig. 1: RoI determination. The bounded red region match to
necessary. This way, we update the image background using the RoI.
the Adaptive Mixture of Gaussians (AMG) [13].
The RoI size was tested in all database. The head region is
B. Moving Objects Classication fully in the RoI in all motorcycle images.
The next step consists of classify the moving objects
resulting from the segmentation procedure. In pattern IV. S UB - WINDOW C ALCULATION
recognition works it is necessary to extract image features at In order to improve further the helmet search we computed a
rst. This way, we use descriptors to extract image features. sub-window, that contains the exact motorcyclist head region.
The descriptor returns a set of values as result. This set This way, the image classication is improved. The reason is
is called feature vector. After this, classiers are used the descriptor will extract features of the head region, only.
to recognize the image. The feature vector is used by the Before computing the descriptor a set of preprocessing is

142
required, all this preprocessing was made in the RoI. At rst,
it is computed the grayscale image.
After the grayscale image be calculated, a average lter
5 5 is applied to reduce the noise. At next, the Otsu [17]
threshold is computed. It is applied in grayscale image to get a
binary image (black and white). Figure 2c shows the result of (a)
this processing. In the binary image is applied the Sobel [12]
algorithm, that returns the image edges, see Figure 2d.
The next step is the CHT (Circular Hough Transform)
calculation to nd circle-like regions. It was used because the
motorcyclist head region has a circle-like shape. The CHT
computation with edges is computationally faster and has a (b)
better precision. The CHT was set with the computed edges
and get the possible circles. One of the parameters required by
the CHT is the circle size fetched. The radius size was half of
the RoI height. That value was chosen because a helmet has
not a size larger than the RoI.
From the found circles a search is made to return the best
circle (circle with more points in the CHT computation). (c)
Figure 2e shows the best circle found in a RoI image. A
strategy that use only geometric information, like CHT, do
not returns good results because the head and the helmet
have similar shape. Therefore, other features are necessary to
separate the problem class. For this, a sub-window from RoI
is computed, that will correspond to the square circumscribed
(d)
in the circle found (see Figure 2f). Figure 2 shows the steps
of the sub-window calculation.
V. F EATURE E XTRACTION
In order to extract image features, Histogram of Oriented
Gradients (HOG) [11], Local Binary Pattern (LBP) [18] and
Wavelet Transform (WT) [19], [20] were used as descriptor. (e)
A. Histogram of Oriented Gradients
The Histogram of Oriented Gradients (HOG) algorithm [11]
is a feature descriptor that calculates an image histogram of
oriented gradients. The nal descriptor is an one-dimensional
array of histograms extracted from the image. The algorithm
is based on the local object shape and appearance, which in (f)
an image can be represented by intensity gradients or edge
Fig. 2: Sub-window calculation. (a) RoI. (b) Grayscale. (c)
directions. The feature extraction can be done without an edge
Binary image. (d) Edge image. (e) Image with the best circle.
position foreknowledge.
(f) Sub-window image.
B. Local Binary Pattern
The Local Binary Pattern (LBP) has a good performance in
many applications, including classication and texture, image
1 tb
recovery and surface inspection. The original LBP [18] labels a,b (t) = ( ), (1)
the image pixels for a threshold neighborhood of 3 3. Each a a
pixel is compared with the central pixel and the result is the were a = 0 is the scale factor, that represents a contraction or
binary number. Figure 3 shows how the LBP is computed. a dilatation in the signal; b  is the translation.
C. Wavelet Transform In a digital image the wavelet coefcients was computed
considering the lines and columns as unidimensional images.
The Wavelet Transform (WT) is an alternative approach in Figure 4 show an image and the four result images after apply
relation to Fourier transform for image processing. The WT the WT.
allows that both frequency and time information are presented.
All pixels correspondent to compress image(Figure 4b,
WT was formally proposed by Deubechies and Mallat [19],
top left image) were used as descriptor in the work. The
[20]. Equation 1 dene the WT,
tree others decompositions was tested too, but they do not

143
amount of motorcyclists with helmet. A CCD video camera
was used. The image database is distributed as follows:
151 images of motorcyclists with helmet; and
104 images of motorcyclists without helmet.
The Figure 5 shows examples of images.

(a) (b) (c)

Fig. 3: Example of LBP descriptor calculation. (a) Window


example. (b) Threshold based on central pixel. (c) Pattern
computed from the threshold result.

Fig. 5: Example of image from database.

IX. R ESULTS AND D ISCUSSION


(a) (b)
The descriptors were applied in the sub-windows. It was
Fig. 4: Result of applying the WT. (a) Original image. (b) tested the combinations as following: WT, HOG, LBP,
Image results. WT+LBP, WT+HOG, HOG+LBP and WT+HOG+LBP. The
signal + is the concatenation between the features vectors,
that we call hybrid descriptors.
return satisfactory results. In this paper was used the Wavelet The HOG descriptor has been set up with a 9 histograms
Transform of Haar. by 9 partitions window, we tested other values but that was
VI. I MAGE C LASSIFICATION the best. This way, a vector of 81 features is generated. Here,
it was also used a large variation of histogram and partition
In this step we use six classiers to compare the results: sizes to conclude which one is the best.
Support Vector Machines (SVM), MultiLayer Perceptron The LBP was used with a window, the image was divided on
(MLP), Radial Basis Function Network (RBFN), Naive Bayes, 9 windows corresponding to 9 histograms. The neighborhood
Random Forest and K-Nearest Neighbors (KNN). These used was 33 for the computation of the label. The labels are
classiers were used because they are the most present in the of 0 to 255. A grayscale image was used for the processing.
literature. A feature vector is generated for each sub-window. In this paper we use the WT of Haar. This transform result
The input of the classiers is the feature vector computed 4 decomposition images. We use the all the pixels of the rst
in the feature extraction step. The images were divided into decomposition (Figure 4) to make the feature vector.
two groups: with-helmet and without-helmet. We was tested the classiers SVM, RBFN, MLP, Naive
VII. M ETHODOLOGY FOR P ERFORMANCE E VALUATION Bayes, Random Forest and KNN. The classiers had the
feature vectors as input. We generated 42 combinations
In order to evaluate the algorithm performance, we used: between descriptors and classiers (7 descriptors and 6
Specicity (S), Positive Predictive Value (PPV), Precision (P), classiers).
Recall (R), F-Measure (FM), Kappa index (K) and Accuracy The SVM classier was performed using linear and
(A). All these measures can be calculated based on four values: polynomial kernel functions. Better results were obtained from
True Positive (TP), False Positive (FP), False Negative (FN) these functions in comparison to radial basis functions and
and True Negative (TN) [21]. tangent sigmoid functions. In relation to the MLP classier,
The level of accuracy of the Kappa index was classied , only one hidden layer with 50 neurons. The RBFN data were
according to the established by Landis and Koch [22]. normalized rst before being trained, is was used two basis
VIII. I MAGE DATABASE functions. The Random Forest classier was performed with
80 trees and an unlimited depth. The Naive Bayes classier
The images were captured in a location where the number
was used with class estimators. The KNN was with K=2, this
of motorcyclists without helmet was balanced regarding the
was the best valeu for K.

144
We use the K-fold cross-validation statistic method with HOG and LBP descriptors. At end, Figure 12 show the ROC
K = 10 to generate the results. In 10-fold cross-validation, curves to WT, HOG and LBP combination.
the original sample is randomly partitioned into 10 equal
size subsamples. Of the 10 subsamples, a single subsample
is retained as the validation data for testing the model, and

the remaining 9 subsamples are used as training data. The
cross-validation process is then repeated 10 times (the folds),
with each of the 10 subsamples used exactly once as the
validation data. The advantage of this method over repeated
random sub-sampling is that all observations are used for
both training and validation, and each observation is used for


validation exactly once. In stratied 10-fold cross-validation,
the folds are selected so that the mean response value is  
approximately equal in all the folds.  
  
Table I shows the best result obtained, that was performed
using the HOG descriptor and the MLP classier, was obtained  !"
  
an accuracy rate of 0.9137. From 255 images, 22 were  #
  $%
incorrectly classied. Therefore, according the Kappa index
the result was Excellent (K 0.8). The FM rate was  !&
  '(
0.9281, that summarizes the P (0.9161) and R (0.9404) rates.  


  '$
TABLE I: The best result for the helmet detection.  )""
  
  
E VPN P R FM K A  
0.8750 0.9100 0.9161 0.9404 0.9281 0.8203 0.9137
Fig. 6: ROC curve using the WT.
Table II shows the obtained results for the SVM classier. It
is observed that the best results was obtained using the hybrid
composition HOG and LBP (accuracy rate of 0.8784). 
The results from the RBFN classier are show in Table III.
The best result was obtained with WT+HOG+LBP (accuracy
rate of 0.8863). In some cases the hybrid descriptors can
improve the results of the individual descriptors.
Table IV show, besides the best result, the results of
another combinations. The Naive Bayes results are presented


in Table V. This classier presented the worst results in


relation to Kappa index, the result was Bad.
 
Table VI show the Random Forest results. The best result
was an accuracy rate of 0.8980. That result was the second  
  
better between all results. At end, the results for the KNN was  !"
  #
show in Table VII.
In order to illustrate the performance of the classier, the  $
  
Receiver Operating Characteristic (ROC), also known as ROC  "!%
  #
curve, was used. The ROC curve was generate for each
combination between descriptor and classier. Figure 6 show  &'(
  )*
the ROC curves and the Area Under Curve (AUC) using the  +""
  ,)
WT. The biggest area was 0.9584 by the classier Random   

Forest. Figure 7 show the ROC curves for the classiers using 
the HOG descriptor. This gure illustrate the curve of the best Fig. 7: ROC curve using the HOG.
result. Figure 8 present the curves of the LBP descriptor. It is
observed that the smallest area between all results is 0.5850,
and is was obtained using the KNN with the LBP descriptor. X. C ONCLUSIONS AND F UTURE W ORKS
The ROC curves using the combination between the WT and This paper comes about the detection of motorcyclists
LBP are shown in Figure 9. The results of the concatenation without helmet. We propose a computer vision system, that
using WT and HOG are shown in Figure 10. The ROC curves is divided as follow: moving objects segmentation, moving
of Figure 11 was obtained by the combination between the objects classication and helmet use detection.

145
TABLE II: Helmet detection using the SVM classier.
Descritor VP FP VN FN E VPN P R FM K A
WT-[1] 137 19 85 14 0.8173 0.8586 0.8782 0.9073 0.8925 0.7301 0.8706
HOG-[2] 139 19 85 12 0.81731 0.8763 0.8797 0.9205 0.8997 0.7456 0.8784
LBP-[3] 133 30 74 18 0.7115 0.8043 0.8159 0.8808 0.8471 0.6032 0.8118
[1]+[3] 137 19 85 14 0.8173 0.8586 0.8782 0.9073 0.8925 0.7301 0.8706
[1]+[2] 137 19 85 14 0.8173 0.8586 0.8782 0.9073 0.8925 0.7301 0.8706
[2]+[3] 139 19 85 12 0.81731 0.8763 0.8797 0.9205 0.8997 0.7456 0.8784
[1]+[2]+[3] 137 19 85 14 0.8173 0.8586 0.8782 0.9073 0.8925 0.7301 0.8706

TABLE III: Helmet detection using the RBFN classier.


Descritor VP FP VN FN E VPN P R FM K A
WT-[1] 131 18 86 20 0.8269 0.8113 0.8792 0.8675 0.8733 0.6924 0.8510
HOG-[2] 133 22 82 18 0.7885 0.8200 0.8581 0.8808 0.8693 0.6733 0.8431
LBP[3] 108 24 80 43 0.7692 0.6504 0.8182 0.7152 0.7633 0.4711 0.7373
[1]+[3] 132 26 78 19 0.7500 0.8041 0.8354 0.8742 0.8544 0.6308 0.8235
[1]+[2] 135 22 82 16 0.7885 0.8367 0.8599 0.8940 0.8766 0.6887 0.8510
[2]+[3] 110 17 87 41 0.8365 0.6797 0.8661 0.7285 0.7914 0.5454 0.7725
[1]+[2]+[3] 138 16 88 13 0.8462 0.8713 0.8961 0.9139 0.9049 0.7635 0.8863

TABLE IV: Helmet detection using the MLP classier.


Descritor VP FP VN FN E VPN P R FM K A
WT-[1] 130 15 89 21 0.8558 0.8091 0.8966 0.8609 0.8784 0.7103 0.8588
HOG-[2] 142 13 91 9 0.8750 0.9100 0.9161 0.9404 0.9281 0.8203 0.9137
LBP[3] 107 32 72 44 0.6923 0.6207 0.7698 0.7086 0.7379 0.3938 0.7020
[1]+[3] 137 16 88 14 0.8462 0.8627 0.8954 0.9073 0.9013 0.7557 0.8824
[1]+[2] 130 16 88 21 0.8462 0.8073 0.8904 0.8609 0.8754 0.7018 0.8549
[2]+[3] 129 21 83 22 0.7981 0.7905 0.8600 0.8543 0.8571 0.6514 0.8314
[1]+[2]+[3] 135 15 89 16 0.8558 0.8476 0.9000 0.8940 0.8970 0.7487 0.8784

TABLE V: Helmet detection using the Naive Bayes classier.


Descritor VP FP VN FN E VPN P R FM K A
WT-[1] 129 18 86 22 0.8269 0.7963 0.8776 0.8543 0.8658 0.6772 0.8431
HOG-[2] 139 21 83 12 0.7981 0.8737 0.8688 0.9205 0.8939 0.7284 0.8706
LBP-[3] 135 85 19 16 0.1827 0.5429 0.6136 0.8940 0.7278 0.0856 0.6039
[1]+[3] 130 18 86 21 0.8269 0.8037 0.8784 0.8609 0.8696 0.6848 0.8471
[1]+[2] 130 18 86 21 0.8269 0.8037 0.8784 0.8609 0.8696 0.6848 0.8471
[2]+[3] 140 51 53 11 0.5096 0.8281 0.7330 0.9272 0.8187 0.4646 0.7569
[1]+[2]+[3] 131 20 84 20 0.8077 0.8077 0.8675 0.8675 0.8675 0.6752 0.8431

TABLE VI: Helmet detection using the Random Forest classier.


Descritor VP FP VN FN E VPN P R FM K A
WT-[1] 142 19 85 9 0.8173 0.9043 0.8820 0.9404 0.9103 0.7692 0.8920
HOG-[2] 144 19 85 7 0.8173 0.9239 0.8834 0.9536 0.9172 0.7850 0.8980
LBP-[3] 139 33 71 12 0.6827 0.8554 0.8081 0.9205 0.8607 0.6228 0.8235
[1]+[3] 141 19 85 10 0.8173 0.8947 0.8812 0.9338 0.9068 0.7613 0.8863
[1]+[2] 144 19 85 7 0.8173 0.9239 0.8834 0.9536 0.9172 0.7850 0.8980
[2]+[3] 144 23 81 7 0.7788 0.9205 0.8623 0.9536 0.9057 0.7505 0.88234
[1]+[2]+[3] 145 19 85 6 0.8173 0.9239 0.8834 0.9536 0.9172 0.7850 0.8980

TABLE VII: Helmet detection using the KNN classier.


Descritor VP FP VN FN E VPN P R FM K A
WT-[1] 112 7 97 39 0.9327 0.7132 0.9412 0.7417 0.8296 0.6436 0.8196
HOG-[2] 119 7 97 32 0.9327 0.7519 0.9444 0.7881 0.8592 0.6948 0.8471
LBP-[3] 64 26 78 87 0.75 0.4727 0.7111 0.4238 0.5311 0.1593 0.5569
[1]+[3] 119 7 97 32 0.9327 0.7519 0.9444 0.7881 0.8592 0.6948 0.8471
[1]+[2] 116 7 97 35 0.9327 0.7348 0.9431 0.7682 0.8467 0.6727 0.8353
[2]+[3] 117 14 90 34 0.8654 0.7258 0.8931 0.7748 0.8298 0.6216 0.8118
[1]+[2]+[3] 121 7 97 30 0.9327 0.7638 0.9453 0.8013 0.8674 0.7096 0.8549

146

 


   
 
  
 
    !"
  
  !
  
 "
  ##
 #
  $
 ! $
  %&'  "!%
  $&&
 ()*
  &  '()
  *
 +!!
    +""
  *
     
  

Fig. 8: ROC curve using the LBP. Fig. 10: ROC curve using the WT+HOG.

 




   
 
    
  
 !"
  #   !
  "#
 $
  %%  $
  %
 "!&
  ##  ! &
  ""
 '()
  *#  '()
  *
 +""
  %  +!!
  ""
     
   

Fig. 9: ROC curve using the WT+LBP. Fig. 11: ROC curve using the HOG+LBP.

147
R EFERENCES

[1] DNIT, Numero de veculos envolvidos por nalidade do veculo,
Departamento Nacional de Infraestrutura de Transportes, Ministerio dos
Transportes, Tech. Rep., 2011.
[2] J. J. Waiselsz, Mapa da violencia 2013: Acidentes de transito e
motocicleta, Centro Brasileiro de Estudos Latino-Americanos, Tech.
Rep., 2013.
[3] S. F. Bureau, Key statistics of thailand, National Statistical Ofce,
Ministry of Information and Communication Technology, Bangkok,


Tailandia, Tech. Rep., 2012.


[4] O. M. da Saude, Global status report on road safety 2013: supporting
a decade of action, Organizaca o Mundial da Saude, Genebra, Suca,
  Tech. Rep., 2013.
 
   [5] NHTSA, National highway trafc safety administration. trafc safety
facts: Research note motorcycle helmet use in 2011 - overall results,
 !"
  #$ no. April 2012, 2011.
[6] MMWR, Helmet use among motorcyclists who died in crashes and
 %
  $$ economic cost savings associated with state motorcycle helmet laws
- united states, 2008-2010. Modidity and Mortality Weekly Report -
 "!&
  '(( MMWR, Tech. Rep. 23, jun 2012.
[7] C.-Y. Wen, S.-H. Chiu, J.-J. Liaw, and C.-P. Lu, The safety helmet
 )*+
  #( detection for atms surveillance system via the modied hough
transform, in IEEE 37th Annual International Carnahan Conference
 ,""
   on Security Technology., 2003, pp. 364369.
  
[8] C.-C. Chiu, M.-Y. Ku, and H.-T. Chen, Motorcycle detection and
tracking system with occlusion segmentation, in WIAMIS 07, USA,
 
2007.
[9] J. Canny, Finding edges and lines in images, Cambridge, MA, USA,
Fig. 12: ROC curve using the WT+HOG+LBP. Tech. Rep., 1983.
[10] J. Chiverton, Helmet presence classication with motorcycle detection
and tracking, IET, vol. 6, no. 3, pp. 259269, 2012.
[11] N. Dalal and B. Triggs, Histograms of oriented gradients for human
detection, in CVPR, 2005, pp. 886893.
The results were satisfactory. The MLP classier using the [12] I. E. Sobel, Camara models and machine perception, Ph.D dissertation,
HOG descriptor achieved the best result, with an accuracy rate Stanford University, Palo Alto, Calif., 1970.
[13] Z. Zivkovic, Improved adaptive gaussian mixture model for background
of 0.9137. We observed that the combinations between CHT, subtraction, in ICPR 2004, vol. 2, aug. 2004, pp. 28 31 Vol.2.
WT, HOG and LBP return satisfactory results too. [14] R. Waranusast, N. Bundon, V. Timtong, C. Tangnoi, and
P. Pattanathaburt, Machine vision techniques for motorcycle safety
Besides the accuracy, we use the metrics Kappa index and helmet detection, in Image and Vision Computing New Zealand
F-Measure. That metrics proved appropriate to evaluate results (IVCNZ), 2013 28th International Conference of, Nov 2013, pp. 3540.
of the proposed system. It is observed that the metrics return [15] E. Fix and J. L. Hodges, Discriminatory analysis, nonparametric
discrimination: Consistency properties, US Air Force School of Aviation
good results only when the classication had good success Medicine, vol. Technical Report 4, no. 3, pp. 477+, Jan. 1951.
rate. That is not observed with the accuracy rate, mainly in [16] R. Silva, K. Aires, T. Santos, K. Lima, R. Veras, and A. Soares,
problems that have an unbalanced number between objects of Automatic detection of motorcycles on public roads, CLEI Electronic
Journal, vol. 17, no. 4, December 2013.
two classes. [17] N. Otsu, A threshold selection method from gray-level histograms,
IEEE Transactions on Systems, Man and Cybernetics, vol. 9, no. 1, pp.
The presented results are promising, but can be improved. 6266, 1979.
An important step for this is the image capture, that needs [18] T. Ojala, M. Pietikinen, and D. Harwood, A comparative study of
a better quality. One of the future works is the licence plate texture measures with classication based on featured distributions,
Pattern Recognition, vol. 26, no. 1, pp. 5159, 1996.
recognize. For this, it is necessary an image with better quality [19] S. G. Mallat, A theory for multiresolution signal decomposition: the
to recognize the characters. wavelet representation, Pattern Analysis and Machine Intelligence,
IEEE Transactions on, vol. 11, no. 7, pp. 674693, 1989.
The used descriptors return a lots of features and this [20] I. Daubechies, Ten lectures on wavelets, 1st ed. Society for Industrial
can difcult the classication. Considering this, the attribute and Applied Mathematics, Jun. 1992.
[21] D. M. W. Powers, Evaluation: From Precision, Recall and F-Factor to
selection is another future work. It is very important to increase ROC, Informedness, Markedness & Correlation, School of Informatics
the success rates. For this reason, an analyse for this kind of and Engineering, Flinders University, Adelaide, Australia, Tech. Rep.
algorithms is required. SIE-07-001, 2007.
[22] L. J. and K. G., The measurement of observer agreement for categorical
The use of Systems of Multiple Classiers (SMCs) will data. Biometrics, vol. 33, no. 1, pp. 159174, 1977.
be made. SMCs with dynamic structures can be used, for [23] S. Haykin, Neural Networks: A Comprehensive Foundation, 2nd ed.
Prentice Hall, 2001.
example, the mixture of experts [23].
Another future work is the detection of motorcycle
passenger. A motorcycle can carry the motorcyclist and a
passenger, but the proposed system do not detect more than
one helmet in an image. Since a passenger without helmet is
a trafc violation.

148

You might also like