Tezislər / Theses

Research Methodology 
Neural Network comes from the base of Deep Learning (DL) which is 
the sub field of Machine Learning (ML). Neural Network uses algorithms 
being inspired by the human brain structure. Firstly, Neural Network is 
presented by data, then it trains itself to recognize the patterns of this data, 
as a result it predicts the output. The training process can reach to even a 
year. Crash analysis method is applied and after a 120 photo training, 
Convolutional Neural Network system can predict, there is a crash exist or 
not with 34% accuracy. 
The figure 1 illustrates the simple working principles of Convolutional 
Neural Network system. The first 3 orange circles are inputs of system, middle 
blue and green circles are hidden layers, and the last red one is output. 

---

THE 3
rd
 INTERNATIONAL SCIENTIFIC CONFERENCES OF STUDENTS AND YOUNG RESEARCHERS 
dedicated to the 99
th
anniversary of the National Leader of Azerbaijan Heydar Aliyev
122
Figure 1.
Convolutional Neural Network system 
Firstly, the image is divided to pixels and the pixels are sent as inputs to 
system, all inputs has a line connected to hidden layer by which are multiplied 
to certain coefficient (between 0-1). This term is again and again will be 
applied to hidden layers and at the end last hidden layers which are 
connected to output will present a final coefficient. By the use of final 
coefficient system decide that there is an accident or not. 
Experimental results 
The structure covered in Research Methodology has been applied to 
practice by the use of Deep Learning Ensemble in Python. In real time video 
is represented to program to analyze if there is an accident or not. Initially, 
the video splits to frames (figure 2), then processed to homogenize the 
images to grayscale and downscale the width to 640 pixels. Then the trained 
CNN model generate a preliminary prediction.