KINETIK, Vol. 3, No. 1, February 2018, Pp.

57-66
ISSN : 2503-2259
E-ISSN : 2503-22677 57
Vehicle Classification using Haar Cascade Classifier Method in
Traffic Surveillance System

Moch Ilham Ramadhani*1, Agus Eko Minarno2, Eko Budi Cahyono*3

1,2,3
Universitas Muhammadiyah Malang
[email protected]*

Abstract
Object detection based on digital image processing on vehicles is very important for
establishing monitoring system or as alternative method to collect statistic data to make efficient
traffic engineering decision. A vehicle counter program based on traffic video feed for specific
type of vehicle using Haar Cascade Classifier was made as the output of this research. Firstly,
Haar-like feature was used to present visual shape of vehicle, and AdaBoost machine learning
algorithm was also employed to make a strong classifier by combining specific classifier into a
cascade filter to quickly remove background regions of an image. At the testing section, the output
was tested over 8 realistic video data and achieved high accuracy. The result was set 1 as the
biggest value for recall and precision, 0.986 as the average value for recall and 0.978 as the
average value for precision.

Keywords: Haar-like Feature, AdaBoost, Cascade Classifier, Vehicle Detection, Digital Image
Processing

1. Introduction
Transportation nowadays is a primary need for every person in finding to most suitable
daily transportation. However, there is an existing huge problem. The uncontrolled personal
vehicle growth has become one of serious transportation problems. According to the previously
conducted research by Indonesia Ministry of Transportation, Indonesian vehicle growth exhibits
surprising results, 12% for motorcycle, 8.89% for car, and 2.2% for bus [1]. In addition, the
research proves that the need of transportation is increasing every year with many factors
triggering this increment exemplified by geography, technology, social and economy [2].
The vehicle growth has increased high traffic density, especially if not being handled
properly. This trend can be the main cause of traffic jam [3]. For example, in Indonesia capital
city, Jakarta has merely 0.01% of road growth. This imbalance vehicle growth increasing 11%
from previous year results bad traffic jam becoming common situation in Jakarta. To solve this
problem, there is an option as building more roads or establishing better traffic engineering [2].
Traffic engineering is an operation technique related with planning, geometric design, and
traffic operation. One of its procedures is measuring PCU (Passenger Car Unit). As conducted by
the previous research [4] to obtain relevant data, researchers manually count passing vehicles
with specific criteria to find out the value of PCU. A manual counting by those researchers has
high probability for any mistakes or human errors caused by, take for example, environmental
conditions or internal disturbances experienced by researchers.
The conventional method to obtain data required big funds and considered as less
efficient. Therefore, it has become a trigger for some research for developing programs intended
for vehicle counting using video feed in obtaining the data. Afterwards, the data will be processed
by using background subtraction method, separating between background and foreground to find
the centroid position inside the frame and counting vehicles passing the counting line [5] and [6].
Unfortunately, the program only counts and assumes centroid as a vehicle without any further
information on the vehicle type. Hence, in order to obtain PCU, the vehicle type should be
recognized because each type of vehicle has different value to count PCU.
In solving this problem, an application to count specific type of vehicle (by focusing only
for cars) using traffic video feed as data, processed by Haar Cascade Classifier method, will be
important. Previously, this method has been also used in the research [7] to detect traffic cones
as obstacles to be avoided by wheeled robot, it was also used in the research [9] to find region of
eyes for making region of interest to capture eye winks as an alternative method to input
Ramadhani, M., Minarno, A., & Cahyono, E. (2017). Vehicle Classification using Haar Cascade
Classifier Method in Traffic Surveillance System. Kinetik : Game Technology, Information System,
Computer Network, Computing, Electronics, and Control, 3(1).
Paper submitted on October 11, 2017; Revision on November 02, 2017; Received December 15, 2017
58 ISSN: 2503-2259; E-ISSN: 2503-2267
password. According to these research [7], [8] Haar Cascade Classifier has been proven to be an
effective and accurate method to detection and specific object recognition.

2. Research Method
2.1 Haar Cascade Classifier
Haar Cascade Classifier is a method utilized for detecting object, also called as Viola
Jones method due to its introduction by Paul Viola and Michael Jones for face detection. This
method has 4 points for detecting an object, such as Haar-like feature, integral image, AdaBoost
learning and Cascade Classifier[9].

Figure 1. Edge Feature [10]

Figure 2. Line Feature [10]

Figure 3. Four-rectangle Feature [10]

Figure 4. Feature Detection Scheme

KINETIK Vol. 3, No. 1, February 2018: 57-66

KINETIK ISSN: 2503-2259; E-ISSN: 2503-2267 59
Haar-like feature is a rectangular feature providing specific indication to an image. Figure
1, Figure 2 and Figure 3 are the examples of common variety of Haar-like feature. Haar-like
feature offers high speed computation depending the number of pixels inside the rectangle feature
and not depending on each pixel value of the image [9]. In obtaining object detection value, Haar-
like feature value was calculated using integral image. Integral image could calculate values
accurately and relatively quick by creating new presentation of image by using value of region
previously scanned by specific Haar-like feature as shown in Figure 4. The value of integral image
was obtained by sum value of previous index, started by left top until right bottom; moreover,
Integral image could be calculated using Equation 1, for example the input and new presentation
depicted in Figure 5(a) and 5(b).

𝑠(𝑥, 𝑦) = 𝑖 (𝑥, 𝑦) + 𝑠(𝑥, 𝑦 − 1) + 𝑠(𝑥 − 1, 𝑦1) + 𝑠(𝑥 − 1, 𝑦 − 1) (1)

a b
Figure 5. (a) Input Image and (b) Integral Image of Input Image

Value which had been calculated by using integral image would then be compared with
the threshold value of specific features provided by AdaBoost. This should be completed to find
potential features because not all features were relevant to use for specific object detection.
AdaBoost combines potential features called weak classifier to become strong classifier. Weak
classifier means less accurate or also irrelevant prediction [11]. Relevan and irrelevan features
shown by Figures 6.

All Features

Relevant Feature Irrelevant Feature

Figure 6. Examples of Relevant and Irrelevant Features

Strong classifier made by AdaBoost can detect object level by level on a cascade. Every
sub-window was scanned for specific criteria on each step; furthermore, a sub-window containing
positive object was used as a feed for the next level filtering with more specific criteria until
obtaining a sub-window which was predicted as a car. On the other hand, a sub-windows not
containing positive object was marked as background and separated by sub-window containing
positive object as shown by Figure 7 [11].

Vehicle Classification using Haar Cascade Classifier Method in Traffic Surveillance System
Moch Ilham Ramadhani, Agus Eko Minarno, Eko Budi Cahyono
60 ISSN: 2503-2259; E-ISSN: 2503-2267

All Vehicle Object

Sub-windows

Sub-windows not containing

positive image marked as
background

Figure 7. Cascade Object Detection Scheme

Figure 8. Region of Interest

2.2 Region of Interest

Region of interest is a specific region of an image that is considered as important in this
study in obtaining data. Haar Cascade Classifier would not need to scan the entire image regions
due to data collection process focused in specific image regions for object detection, object
tracking and object counting, illustrated in Figure 8.

2.3 Object Tracking

Object tracking was used to obtain the specific position (x,y) of object inside the frame to
be compared with list of previous positions of tracked objects; however, new positions or positions
not including on the list of tracked object positions was added as a position (x,y) of a new object.
If the new position was included in the list of positions of previous tracked objects, it would be
used as a new position of a recognized object. The general process of object tracking is presented
by the flowchart in Figure 9.

2.4 Object Counting

Every passing vehicle object inside ROI (Region of Interest) was tracked based on its
position and would be compared with the list of tracked object positions. For a new position or
position not including in the list of tracked objects, it was be added as a new object and should
be counted. If the new position was included in the list of positions of previous tracked objects, it
means the position had already been counted as a recognized vehicle. The general process of
object counting will be illustrated by the flowchart in Figure 10.

KINETIK Vol. 3, No. 1, February 2018: 57-66

KINETIK ISSN: 2503-2259; E-ISSN: 2503-2267 61

Figure 9. Flowchart of Object Tracking

Figure 10. Flowchart of Object Counting

Vehicle Classification using Haar Cascade Classifier Method in Traffic Surveillance System
Moch Ilham Ramadhani, Agus Eko Minarno, Eko Budi Cahyono
62 ISSN: 2503-2259; E-ISSN: 2503-2267
3. Results and Discussion
This section will introduce the result of the vehicle detection efforts. A total of 1000 vehicle
images presented in Figure 11 and 200 negative images were used in the training procedure used
OpenCV library. Positive samples were obtained by manually cropping and resizing to 30 x 30
pixels. Moreover, the vehicle area of the vehicle images was recorded with traffic surveillance
camera depicted in Figure 11. The testing phase employed 8 videos as realistic data test of traffic
surveillance video as shown in Table 1. Each video was taken using 30 o – 50o camera position
facing down the road as illustrated in Figure 12(a), and Figure 12(b) presents the video footage.

Figure 11. Positive Image Examples

a b
Figure 12. (a) Camera Position and (b) Video Result of Camera Position Perseption

Table 1. Video Test Description

No Video Video Capture Video Shake Car
Test Time Stabilizer Level Total
1. 1.mp4 Afternoon Tripods Rare 41
2. 2.mp4 Evening Hand Low 71
3. 3.mp4 Morning Tripods Rare 50
4. 4.mp4 Evening Hand Low 69
5. 5.mp4 Afternoon Tripods Rare 41
6. 6.mp4 Evening Hand Low 86
7. 7.mp4 Afternoon Tripods Rare 60
8. 8.mp4 Afternoon Tripods Low 44

3.1 Detection Rate Testing

The detection rate testing measures the detection rate of classifier to obtain recall and
precision value of classifier with different scale factor value to determine its value giving the best
performance to the classifier.

KINETIK Vol. 3, No. 1, February 2018: 57-66

KINETIK ISSN: 2503-2259; E-ISSN: 2503-2267 63
Table 2. Detection Rate Testing
Video Scale Actual TP FP FN Recall Precision
Test Factor Total 𝑇𝑃 𝑇𝑃
( ) ( )
Vehicle 𝑇𝑃 + 𝐹𝑁 𝑇𝑃 + 𝐹𝑃

1.mp4 1.3 41 40 1 0 1 0.975

2.mp4 1.3 71 69 2 0 1 0.971
3.mp4 1.3 50 49 1 0 1 0.980
Average (Scale factor 1.3) 1 0.975
1.mp4 1.5 41 40 1 0 1 0.975
2.mp4 1.5 71 63 8 0 1 0.887
3.mp4 1.5 50 47 4 0 1 0.921
Average (Scale factor 1.5) 1 0.927
1.mp4 1.5 41 40 1 0 1 0.975
2.mp4 1.5 71 67 4 2 0.97 0.943
3.mp4 1.5 50 47 4 0 1 0.921
Average (Scale factor 1.7) 0.99 0.946
1.mp4 1.7 41 38 3 0 0 1
2.mp4 1.7 71 2 69 0 0 1
3.mp4 1.8 50 44 7 0 0 1
Average (Scale factor 1.9) 1 0.593

For data test as shown in Table 2, the classifier was tested using 3 samples of video test
and 4 different values of scale factor. Each scale factor value resulted different detection rate. To
archive high detection rate, the scale factor value should be determined giving the best
performance to the classifier. According Table 2, the “1.3” of scale factor value giving the best
performance for the classifier would be used in vehicle counting test section; furthermore, the
average value of recall was set as “1”, and the average precision value was “0.975”.

3.2 Vehicle Counting Test

This section will be testing section for counting car as a final result of following functions such
as detecting, tracking and counting.

Table 3. Vehicle Counting Result

No Video TP FP FN Recall Precision
Test
1 1.mp4 40 1 0 1 0.975
2 2.mp4 69 2 0 1 0.971
3 3.mp4 49 1 0 1 0.980
4 4.mp4 67 1 3 0.957 0.985
5 5.mp4 41 0 1 0.976 1
6 6.mp4 82 4 4 0.953 0.953
7 7.mp4 59 1 0 1 0.983
8 8.mp4 43 1 0 1 0.977
Average 0.985 0.978

As informed in Table 3, Haar Cascade Classifier was tested using 8 traffic videos and
resulted “1” as maximum value for precision and recall, having the average recall set as “0.986”
and average precision set as “0.978”.

4. Conclusion
This research designs a classification system to determine object as specific type of
vehicle. Haar Cascade Classifier proposed by Paul Viola and Michael Jones is used to determine
object as car and counted the number of passing vehicles on the specific road using traffic videos
as input. The detection rate of this system is affected by the scale factor value, different scale
factor value providing varied detection rates. In obtaining high detection rate, the scale factor
value giving the best performance to classifier should be determined. In the future, providing

Vehicle Classification using Haar Cascade Classifier Method in Traffic Surveillance System
Moch Ilham Ramadhani, Agus Eko Minarno, Eko Budi Cahyono
64 ISSN: 2503-2259; E-ISSN: 2503-2267
accurate and robust vehicle detection system will still be a challenging task in the field of intelligent
transportation surveillance systems. The researchers will conduct further research for comparing
background subtraction and Haar Cascade Classifier method or combining background
subtraction and Haar Cascade Classifier to detect more specific vehicle types.

5. Notation
s(x,y) : Representation of Integral Image (x,y)
i(x,y) : Current Representation of Image (x,y)
Positive Image : Recognized Object
Negative Image : Not Recognized Object or Background Image
TP : True Positive or Detected Vehicle
FP : False Positive or Undetected Vehicle
FN : False Negative or Missed Classification/Missed Detection

References
[1] Ministry of Transportation, "Transportation Statistics," Vol. l, 2016.
[2] Central Bureau of Jakarta Province, "Transportation Statistics of Jakarta 2015," 2015.
[3] N. Redhantika, “Traffic Density in Malang City," Universitas Merdeka Malang, Pp. 1–10,
2014.
[4] R. Anwar, “Measuring Passenger Car Unit in Banjarmasin Municipality," Vol. 1, No. 1, Pp.
22–27, 2000.
[5] Fajar Mit Cahyana, “Design Counting Program on the Number of Vehicles on Uni-directional
Traffic Using C++ Programming Language with OpenCV Database," Universitas Brawijaya,
2014.
[6] A. Helmi, “Application of Traffic Density Level Detection Based on the Number of Passing
Vehicles with OpenCV,” 2015.
[7] C.-J. Lee, “Obstacle Detection and Avoidance Via Cascade Classifier for Wheeled Mobile
Robot,” Int. Conf. Mach. Learn. Cybern., Pp. 5, 2015.
[8] M. Syarif, P. Studi, T. Informatika, F. I. Komputer, U. Dian, and N. Semarang, “Blink Detection
with Haar Cascade Classifier and Contour for Password Login,” Techno.com, Vol. 14, No. 4,
Pp. 242–249, 2015.
[9] P. Viola and M. M. J. Jones, “Robust Real-time Face Detection,” Int. J. Comput. Vis., Vol.
57, No. 2, Pp. 137–154, 2004.
[10] A. Mordvintsev, “OpenCV-Python Tutorials Documentation Release Beta,” 2017.
[11] J. Howse, "OpenCV Computer Vision with Python," 2013.

KINETIK Vol. 3, No. 1, February 2018: 57-66

KINETIK ISSN: 2503-2259; E-ISSN: 2503-2267 65

Vehicle Classification using Haar Cascade Classifier Method in Traffic Surveillance System
Moch Ilham Ramadhani, Agus Eko Minarno, Eko Budi Cahyono