Page 67 - 2023-Vol19-Issue2

63 |
II. RELATED WORK
In this section, the main previous works that address the ob-
ject detection process using deep learning models are covered.
Ayegu¨l Uc¸ar et al. [11] suggest two Convolutional Neural
Networks (CNNs) deep models with different layers and use
linear Support Vector Machine (SVM) classifiers in the train-
ing features step. This work uses Caltech 101 and Caltech
pedestrian detection datasets. The average accuracy rate in
this work is 92.80+0.43%.Kai Janet et al. [12] utilize faster
region Convolutional Neural Networks (faster RCNN) with
Residual Neural Networks (ResNet 101) as adeep model. Use
it to recognize the scratches points on the wheel hub even
if the image data has a complicated background, in addition,
the model can recognize variance types of wheel hub defects
as well as determine the class and position of the defective
area. This work establishes a database consisting of defect
area images and achieves an average accuracy rate of 86.3%.
Stefan Schneider et al. [13] compare two deep learn-
ing models used for object detection namely, Faster Region-
Convolutional Neural Network together with You-Only-Look-
Once v2.0 (YOLOv2) in the Serengeti dataset, the results
prove that the first method explicitly promises accuracy of
93%, where YOLO fails to perform the same accuracy and
achieves 76.7%.
Chandan G et al. [[14] develop a MobileNets-based tech-
nique with a Single-Shot Detector (SSD) technique for track-
ing as well as detection of the camera-based dataset in a
python environment. These techniques include detecting the
region of interest of an object from a given image class and
achieving an average accuracy rate of 99.0%. CNN is used
for feature extractions.
Xiangmo Zhao et al. [15] propose a model that uses the
complementarity of the camera and 3D LIDAR and camera
data to recognize multiple objects around an autonomous car.
The average accuracy rates of the accuracy of the proposed
method reached 89.04% and 78.18%, respectively in the de-
tection of vehicles and pedestrians at a moderate level of
difficulty.
Muhammad Shakeel et al [16] propose an accurate drowsi-
ness detection methodology based on object detection using
Convolutional Neural Networks. Which achieves an accuracy
of 84% in their own new, annotated Drowsy dataset.
Muhammad Yahya et al. [17] provide an improvement
in object detection using LiDAR-collected data by using dif-
ferent algorithms like YOLOv2, and Interactive Multi-Mode
(IMM). The average accuracy rates in the newly collected data
set are 71.43% and 57.14% respectively.
YANFEN LI et al [18] suggest an object recognition as
well as detection theme for autonomous driving. The sug-
gested theme is capable to detect ten kinds of objects; the
proposed model is based on YOLOv4 and achieves an aver-
Odey & Marhoon
age accuracy rate of 52.7% in BDD100k and three collected
datasets.
In this work feature extraction and feature expansion of
the udacity images dataset are performed to assist deep model
classification the features are extracted by a linear discrimi-
nant analysis and the feature expansion contribute six opera-
tions namely: standard deviation, min, max, mod, variance,
and mean classified as five classes which are bikers, cars,
pedestrians, traffic lights and trucks.
III. METHODOLOGY
the proposed method is based on five main stages, as ex-
plained in Figure-1: the first stage, is the data collection stage
where we use data from the Udacity Self Driving Car Dataset
dataset. Our data is distributed in five classes,namely: bakers,
cars, pedestrians, traffic lights, and trucks, each class has a
different number of files.
The second stage is an image augmentation step, the ex-
pected result from this stage is more data where we will get
images for objects with different angles and positions as well
as increase the number of images in each class so that the
number of images in the different classes will be quit equally
to overcome the over-fitting issues.
On the other hand, the third stage, the pre-processing stage
in which we get a more enhanced dataset where we manip-
ulate the smoothing and blurring of our images to make the
detection of the desired objects easier and possible under all
conditions of noise and lightness issues so that the dataset
will be more appropriate to the next feature extraction stage
whereas, the fourth stage is the feature extraction and expan-
sion stage where we extract features from our image dataset,
then feed these features to our hybrid model. The final stage is
our proposed supervised deep learning model which employs
the benefits of CNN and Dense techniques to implement a
hybrid model that mimics the human brain in case of learning
based on features instead of images used to achieve object
detection dedicated to self-driving cars.
A. The Dataset
This study was evaluated based on features extracted from
udacity dataset, which is dedicated mainly to data to be pro-
vided in autonomous cars. Primarily 110 MB dataset was
downloaded, then many processes are accomplished to en-
hance and improve our dataset before turning them into fea-
tures, through the data augmentation and image processing
steps.
B. Data Augmentation
Data augmentation procedures apply different operations
to the images data to provide more images for objects with
different angles and positions as well as increase the number