IRJET - Traffic Density Estimation by Counting Vehicles using Aggregate Channel Features

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 07 Issue: 07 | July 2020 www.irjet.net p-ISSN: 2395-0072
© 2020, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 3424
TRAFFIC DENSITY ESTIMATION BY COUNTING VEHICLES USING
AGGREGATE CHANNEL FEATURES
Joel Shaji1, Anisha Mohammed2
1Scholar, Dept. of Electronics and Communication Engineering, College of Engineering Kallooppara,
Kerala, India
2Asst. Professor, Dept. of Electronics and Communication Engineering, College of Engineering Kallooppara,
Kerala, India
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - Vehicle detection methods have become useful
for many artificial intelligent systems such as self-driving
cars or autonomous vehicles and traffic congestion control
methods. The most popular Google car are based on the
result of vehicle detection methods from the driver
perspective. The vehicle detection methods have huge
amount of benefits in the modern world. Vehicle detection
methods can be classified on the basis of basic methods and
computer vision based methods. The primitive methods
include the basic image processing techniques to detect the
vehicle from its background. A comparative study along
with a new method using aggregate channel feature is used
to determine the density estimation of vehicle. The results
can used for setting the flow of traffic lights. For this
comparative study, MATLAB have been used. Every method
from image preprocessing to object detection were utilized
through image processing and computer vision.
Key Words: ROI, Threshold, ACF, BLOBs, Computer Vision,
Morphology
1. INTRODUCTION
Vehicle congestion is becoming more popular in
developed and developing countries as the number of
vehicles in the road increases day by day. Proper
congestion control systems can be introduced to limit the
congestion to an extent, one way of doing it is based on
introducing manual systems such as traffic police to guide
the moment of vehicles in traffic. There are limitation for
such kind of systems as the amount of congestion
increases daily. If proper congestion controlling systems
based on the latest image processing techniques and
computer vision based techniques that rely on the
convolutional neural networks are made along to assist in
the manual method the traffic congestion control will
become more efficient.
The novel method that is introduced here is based on an
evolved form of Viola Jones object detection known as
(ACF) Aggregate Channel Features .The method is more
efficient than the Viola Jones method. The novel method
only needs small amount of data set also the amount of
false positives will be 1000 fold less than the Viola Jones
method for object detection.
The primitive methods as described earlier uses the
normal image processing techniques such as threshold,
morphology and Edge detection method. The commonly
used method for object detection includes the BLOB
analysis which identifies the region of interest based on
certain parameters. When including BLOB analysis with
the primitive methods that were used the vehicles can be
detected. The drawbacks of the primitive methods used is
that the image for detection should contain the foreground
(vehicle) and background that can be well distinguished
and these methods cannot be used for high level of traffic
congestion.
The modern computer vision based method such as the
RCNN or the Fast RCNN methods seems to be more
reliable in identifying object in an image. The novel
method used can be incorporated with the Fast RCNN
method to make the object detection real time and
accurate. Further computer vision methods such as the
newly introduced method YOLO (You Only Look Once) can
be introduced along with the novel method. The YOLO
method is a multiple object detection method which helps
in identifying one or more objects in the image.
2. LITERATURE SURVEY
The simplest method for vehicle detection was introduced
by [1] which describes the basic methods such as
threshold and morphology that can be used to detect
vehicles from a well distinguished image. It is one of the
simplest method for calculating the density of vehicles in
an image. The method have less accuracy in distant frame
videos. Real time density estimate method [2] which
selects the region of interest by background subtraction

for detecting the vehicle from the image. Gaussian mixture
model is used for ROI based method, the BLOBs analysis is
incorporated along with background subtraction. Gaussian
mixture method with the Kalman filter [3] was used for
multiple vehicle detection based on the shadow removal
technique. Another approach based on the Kalman filter
tracking [4] is based on the color feature of the vehicle. For
more accuracy in finding the vehicle a method based on
Canny edge detection was introduced [5] the paper
implements Canny edge detection along with the basic
BLOBs analysis method for finding the vehicle. A similar
approach was introduced in [6] which uses the same
principle of edge detection but uses Gabor filter for the
precision in detection. [7] Describes the Linear Quadratic
Estimation method which uses mask and morphological
operations to detect the vehicle.
The computer vision based method [8] which used the
Viola Jones algorithm for object detection. The method can
also be described as cascade object detector it has more
accuracy in finding the vehicles when compared to the
previous methods described. The drawback of this method
is that the aspect ratio of object that are similar are only
identified by the detector. The aggregate channel feature
method [9] similar to the cascade object detector was
introduced for detecting the face but unlike the cascade
object detector it uses the sliding window approach and
has more accuracy in finding multiple objects in an image.
Fast RCNN method [10] uses the region based
convolutional network for the detection of objects, the
method is an evolved form of the Viola Jones method
incorporating the neural network.
3. IMAGE ACQUISITION
Since real time video is used for the processing the images
are obtained from video frames converted in specific time
period. 24-30 frames are there in a second of video. The
frames which are close to each other are almost similar
therefore we need to remove the frames similar to each
other. For removing the similar video frames, the frames
are obtained in an interval of 5 × 𝑖 𝑡ℎ frame per second.
𝑡 ∑ 𝑖
(1)
If the frame rate of the video is 24 frames per second then
n = 8 and if it is 30 frames per second then n = 10. For the
proposed project 360 frames were obtained in a minute.
For videos with more length the number of images
obtained will increase. The next phase is image
preprocessing for training the detector the images should
be in specified dimensions and this is done using MATLAB
toolbox.
4. IMAGE PREPROCESSING
For training, testing and validating, the images obtained
are shuffled into training data, test data and validating
data. For the proposed system three videos with different
backgrounds were combined and were processed. The
three minute video produced almost 1,080 video frames
and were split into three datasets. Each dataset contain
360 frames for testing, training and validating. But most of
the images are of a certain time frame of similar objects the
number of training images used should be further reduced.
Once images are shuffled the next step is to resize the
images to a specified size so that the training time can be
reduced. When reducing the size of images it shouldn’t be
too small, so that it produces more false positives while
training. For the proposed method the size of resized
images were 320x640x3. This is because we will be
dividing the image into 8x8 and 16x16 patches to extract
the features. For the extraction of features we will be using
the fast pyramid approach based on [11] which uses the
features from different channels of color space. We need to
convert the RGB color space to CIE XYZ and then using
these parameters we can find the CIE LUV color space. If
the chromaticity coordinates of an RGB system ( , ),
( , ) and ( , ) and its reference white ( , , ),
here is the method to compute the 3 × 3 matrix for
converting RGB to XYZ:
[ ] = [ ] [ ] (2)
[M] = [ ] (3)
(4)
(5)
(6)
(7)
=1 (8)
(9)
(10)

(11)
(12)
Once the RGB color space is converted to XYZ color space
using the above parameters we can find the LUV color
space using the following equation:
{
√ 𝑖
(13)
(14)
(15)
Where,
(16)
(17)
(18)
(19)
(20)
ϵ={
𝑡 𝑡
𝑡 𝑡 𝑡ℎ 𝑡
κ={
𝑡 𝑡
𝑡 𝑡 𝑡ℎ 𝑡
5. METHODOLOGY
Now we have all the sources to make a vehicle detector.
The next step is to extract all the features from the
channels found previously. Once the feature pyramid is
formed using the feature descriptor using the multiple
channels which were formed from RGB color space we
need to train the detector using these features. There are
many methods that can be used as a feature measure like
HOG, Haar, SIFT and SURF, but the method that we are
using for this detector is HOG descriptor. Since we need to
detect the vehicles which have similar shape and size from
the image the HOG descriptor is best suited to extract
features based on structure and shape. As in normal cases
we need to find the directional gradient and . Then
the angle or orientation using the Pythagoras theorem
using the Histogram of Oriented Gradient method. As the
name suggests we will make a histogram from the
orientation bins and directional gradients.
Fig -1 Block Diagram of proposed method.
Detecting objects of different scales are difficult so as
proposed by [11] The ACFs have properties of extracting
unique features from the image. For extract the features
the input image will be created as a multiple resolution
pyramid. If an input image is given we compute several
channels it sums every patches of pixels in
image , and the algorithm smooth the resulting lower
resolution channels. The features obtained will be simple
extracted pixel based features as described in the previous
section. [12] Describes different channels from which
features can be extracted that are useful for vehicle
detection the same channels are used here for vehicle
detection. As described the histogram of oriented
gradients is included with it. Normalized gradient
magnitude, histogram of oriented gradients (6 channels)
and LUV color channels are the channels used.
5.1 Adaptive Boosting
For vehicle detection AdaBoost algorithm is used to create
a strong classifier. The Adaptive Boosting consist of stages
were each stage is a group of weak learners. Each of these
stages are trained using a technique called boosting.
Boosting helps in creating a highly accurate classifier by
taking a weighted sum of decisions made by the weak
learners. The steps for the generalized adaptive boosting
algorithm is as follows:
Step 1: Given example images ( , ), … , ( , ) where
= -1, 1 for negative and positive examples.
Step 2: Initialize weights = 1/2M, 1/2L for = -1, 1
respectively, where M and L are the number of negatives
and positives respectively.
Step 3: For t = 1 to T
1) ∑
2) Choose the classifier, ℎ , with the lowest error :
∑ 𝑖 |ℎ
3) Update the weights for each example:

{
𝑖 𝑖 𝑖 𝑡
𝑡ℎ 𝑖
Where βt = εt/(1 - εt t = - log βt
4) The final strong classifier is:
ℎ
{
∑ ℎ ∑ ℎ
𝑡ℎ 𝑖
Pseudo code for the Adaboost algorithm is given above. It
can be defined for the proposed method. First we need to
define the learning data N with images that contain vehicle
as M and the images that does not contain vehicle as L
each item of ( where is
the class to which is defined. The next step is to
initialize sample weights to create the first weak learners.
Third step is the training stage were we need to find the
error rate between predicted values and the previously
available values. The error rate can be defined as the sum
of learning data values which were mistakenly classified
multiplied with the weight values. The weights needs to be
updated until we get a weak classifier with minimum error
rate. The next step is to create a strong classifier as the
sum of weak classifiers. Using a sliding window approach
the detector is made to find whether the vehicle is found in
the region or not. A bounding box is made around the
region of interest if the vehicle is found,
5.2 Correcting Bounding Box
Since our proposed method is for calculating the number
of vehicles in an image and use the result for calculating
the density we need bounding boxes over the region of
interest that was found by the detector. Now there will
be multiple number of bounding boxes over the region
of interest which will be over lapped with each other.
We need to remove these overlapped bounding boxes
and only need a single bounding box over the region of
interest. The overlapped boxes are found using the
criteria:
𝑖
(21)
After the minimum overlapping is found the bounding
boxes with the least confidence can be removed from the
region of interest. The algorithm check for overlapping
and if found the next step is to check for the score or the
confidence score of the bounding box. If the score of box
is higher than the score of the other then the box with
highest score will be maintained and the other will be
removed. The process continues to check for the next
minimum overlap until there will be no more overlaps
present.
Fig -2 Block Diagram for correcting bounding box.
The bounding boxes will be selected according to the
above mentioned block diagram.
4.3 Density Estimation
Once the bounding boxes are created the density of the
vehicles need to be calculated this is done by the counting
of bounding boxes and dividing it with the maximum
density of vehicle that could occur.
𝑖𝑡 ℎ𝑖
(22) Were, 𝑉𝑖 is the total number of bounding boxes with
ROI that can occur. (Maximum number of vehicles that can
accommodate in the area).
Once the density is calculated we can use the data given in
the table below to set the traffic light flow. The table given
(Table 1) is only an assumption that the data can be used to
create further algorithms. The given data only shows how
the timing of the green signal light can be controlled
according to the corresponding density estimates during
peak hours and normal hours of traffic.

Table -1 Setting of traffic light on density.
6. RESULTS AND DISCUSSION
To validate the proposed method, experimental images
were obtained from video camera in real-world road
conditions at various times. The images obtained for the
experiment were 1080 × 920 pixels image. The
experiment was done on a PC with Windows 10 OS (3.3
GHz Dual Core, 4 GB RAM, Single CPU) and MATLAB
2018 a. To reduce the processing time, the size of the
input image was reduced to 480 × 720 pixels by the
bilinear interpolation method. In the proposed method,
the number of positive vehicle images used for training
of the ACF detector was 360 and non-vehicle images
were selected for the remaining area that excluded the
vehicle regions. The average size of the training vehicle
regions for the vehicle detection was 31 × 34, the
repetition period (T) of the ACF detector algorithm is set
to 4, the number of training samples in the non-vehicle
(negative) for each learning step was 4 (S: 4), and the
maximum number of weak classifiers was set to 744.
Fig -4 Sample images used for training.
Before going to the results of the proposed method, we
need to discuss why this method was proposed on other
simplest methods which are available at present. The
first method that was implemented was using the region
of interest based method which highlights the vehicle
from its background using the difference between the
reference image and the target image as shown in figure
(3). Then uses bounding boxes to calculate the number
of vehicles.
Fig -3 Result of using first method.
The result was not even close to the actual number of
vehicles in the image. The bounding boxes will consider
the closely placed objects as a single object. Therefore
the result will not be accurate. This method can be used
for implementing simple density estimation tasks were
the background and foreground images can be well
distinguished.
The next method that was studied was based on basic
image processing techniques like the morphological
Density (%) Time(Sec)
Peak Hour Non-Peak
Hour
0-15 60 s RED 50 s RED
15-25 30 s GREEN 20 s GREEN
75-85 90 s GREEN
>85 120 s GREEN

operations and threshold as in figure (5). The result was
calculated using BLOBs analysis. Like the previous
method the accuracy is not much, since the background
and foreground images are not well distinguished.
The third method was based on Canny edge detection
and the image filling method to create a structure and
then using the BLOBs analysis the number of vehicle is
calculated as in figure (7). The density is estimated as
the total number of vehicles divided by the area of the
image. When compared to the previous two methods the
edge detection method gave better result in detecting
the vehicles.
Now the cascade detector uses the same principle of
finding the gradient as feature like the edge detector. A
cascade vehicle detector was implemented to find its
performance on a single class ‘car’ as shown in figure
(8).
Fig -5 Result of method 2.
Fig -6 Density estimation result of method 2.
Fig -7 Result of method 3.
Fig -8 Result of Cascade Detector.
Finally the result for the proposed method is shown in
figure. Three classes of vehicles class 1 ‘car’, class 2 ‘Auto
R’ and class 3 as ‘Bikes’ were detected using the ACF based
detector. For this there ACF detectors were formed for
each class of vehicle and a single input image was given to
the three detectors which gives results of three classes in
the image.
Fig -9 Result of the proposed method.

Fig -10 Average precision calculated using precision recall
curve.
Table -2 Result of first two test images on proposed
method
Fig -11 Density estimation result of proposed method
The accuracy of the proposed method was calculated using
the average precision from the precision, recall curve.
Where precision is the rate of accurate detection of the
vehicle and recall is the rate of deviation in the detection
area.
The average precision was calculated to be 87%. This was
calculated from the average of the three detectors that
were used to detect the three classes of vehicle. Accuracy
shows how often the vehicle detector provides correct
results.
In case of the cascade object detector, the detector was set
to a training size of [32 , 38] and almost all the positive
samples (825– 859) were used for training. The number of
negative samples were set between the limits (1000 –
1500). The false alarm rate was set as 0.1, the sampling
factor to 2 and the learning cycle to six. From the
experimental result if the learning cycle is increased, the
result will be much better but for this we need to feed the
trainer with more images.
Table- 3 Average precision calculated at different time
7. CONCLUSIONS
Real-time vehicle detector and vehicle density estimator is
proposed. The system uses an approach based on
aggregate channel features. The proposed method
estimates the class of vehicle from the video camera
installed at different traffic junctions. In the proposed
method, the ACF-based AdaBoost algorithm was used to
detect the vehicle region and estimate the density of the
vehicles. In the experiments on various road
environments, the accuracy of the vehicle detection was
estimated to be 87.5%, and the time required for the
processing was 0.76 s per frame.
Also a comparison between the cascade detector and the
proposed method was validated the proposed method
gives an accuracy of 4 percent and also different classes of
vehicles can be detected using a single detector. The
resultant counting of vehicles were used to estimate the
density of the vehicles during peak and non- peak traffic
times. Hence the time for the traffic signals can be
implemented using the method, also this detection method
Test
Image/Dete
cted
Numbers
No of
vehicle
class 1
Car
No of
Vehicle
class 2 Auto
R
No of
vehicle
class 3
Bike
Total
Number
of
vehicles
Test Image
1
12 2 4 18
Detected
number of
vehicle
11 2 3 16
Test Image
2
8 4 5 17
Detected
number of
vehicle
8 3 4 15
ParameterM
easures
Average
Precision
Error rate
(𝛼𝑡)
Training
Time
T = 2, S = 2 0.7847 0.4345 283.57
T = 2, S = 4 0.5877 0.6366 432.75
T = 4, S = 2 0.8593 0.4637 592.87
T = 4, S = 4 0.8767 0.3017 680.17

can be used with the RCNN (Region based convolutional
neural network) to detect and track vehicles for A.I self-
driving vehicles.
The drawback of this method is that the number of
datasets that can be used for training is less, since the
frames are selected from the videos for training. In the
experimental road environment, there was a problem of
misdetection when the driving vehicles overlapped with
the shadows of trees, traffic signs, streetlights or were
represented in a color similar to the background. Also the
problem of detecting different class of object separately
can be solved using further research. Therefore, future
research will focus on improving the accuracy of vehicle
detection and reducing the processing time
REFERENCES
[1] Singh, Abhijeet, Abhijeet Kumar, and R. H. Goudar.
"Online traffic density estimation and vehicle
classification management system." Indian Journal of
Science and Technology 7.4 (2014): 508.M. Young, The
Technical Writer’s Handbook. Mill Valley, CA:
University Science, 1989.
[2] Abbas, Naeem, Muhammad Tayyab, and M. Tahir
Qadri. "Real time traffic density count using image
processing." International Journal of Computer
Applications 83.9 (2013): 16-19.
[3] Jin, Sheng, et al. "Short-term traffic safety forecasting
using Gaussian mixture model and Kalman filter."
Journal of Zhejiang University SCIENCE A 14.4 (2013):
231-243.
[4] Xie, Lei, et al. "Real-time vehicles tracking based on
Kalman filter in a video-based ITS." Proceedings. 2005
International Conference on Communications, Circuits
and Systems, 2005.. Vol. 2. IEEE, 2005.
[5] Balu, Shibin, and C. Priyadharsini. "Smart Traffic
Congestion Control System." 2019 3rd International
Conference on Computing Methodologies and
Communication (ICCMC). IEEE, 2019.
[6] Sun, Zehang, George Bebis, and Ronald Miller. "On-
road vehicle detection using evolutionary Gabor filter
optimization." IEEE Transactions on Intelligent
Transportation Systems 6.2 (2005): 125-137.
[7] Chauhan, Naresh Singh, et al. "Vehicle detection,
tracking and counting using linear quadratic
estimation technique." 2018 2nd International
Conference on Inventive Systems and Control (ICISC).
IEEE, 2018.
[8] Viola, Paul, and Michael Jones. "Rapid object detection
using a boosted cascade of simple features."
Proceedings of the 2001 IEEE computer society
conference on computer vision and pattern
recognition. CVPR 2001. Vol. 1. IEEE, 2001.
[9] Alionte, Elena, and Corneliu Lazar. "A practical
implementation of face detection by using Matlab
cascade object detector." 2015 19th International
Conference on System Theory, Control and Computing
(ICSTCC). IEEE, 2015.
[10] Girshick, Ross, et al. "Region-based convolutional
networks for accurate object detection and
segmentation." IEEE transactions on pattern analysis
and machine intelligence 38.1 (2015): 142-158.
[11] Dollár, Piotr, et al. "Fast feature pyramids for object
detection." IEEE transactions on pattern analysis and
machine intelligence 36.8 (2014): 1532-1545.
[12] Arunmozhi, Ashwin, and Jungme Park. "Comparison of
HOG, LBP and Haar-like features for on-road vehicle
detection." 2018 IEEE International Conference on
Electro/Information Technology (EIT). IEEE, 2018.

IRJET - Traffic Density Estimation by Counting Vehicles using Aggregate Channel Features

More Related Content

What's hot (20)

Similar to IRJET - Traffic Density Estimation by Counting Vehicles using Aggregate Channel Features (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET - Traffic Density Estimation by Counting Vehicles using Aggregate Channel Features