Road detection by image processing, using neural network

Subject Areas : Journal of Simulation and Analysis of Novel Technologies in Mechanical Engineering

Ahmad Keshavarzi ¹ , Mehdi Keshavarz ² , Alireza Moradi ³

1 - Master of Science (MSc), Faculty of Mechanical Engineering, Islamic Azad University, Khomeini Shahr, Isfahan, Iran.
2 - Department of Mechanical Engineering, Khomeinishahr Branch, Islamic Azad University, Khomeinishahr, Iran
3 - Department of Mechanical Engineering, Khomeinishahr Branch, Islamic Azad University, Khomeinishahr, Iran

Received: 2022-06-13 Accepted : 2022-11-06 Published : 2022-09-01

Keywords: image processing, Neural network, Road detection, Automatic driver,

Abstract :

One of the methods in road detection is image processing, which is detected according to the contents of the image captured by the car camera. Much research has been done in this area by researchers, who have used a variety of methods to better process the road and detect the direction of vehicle movement. Using the photos taken by the camera installed on the car and their preprocessing, such as removing inefficient corners and turning it into a gray photo during various stages and applying a goblin filter on the photos and extracting the tallest diagram in order to detect the direction of the road. In this research, standard data processing such as line slope for neural network input and LVQ neural network are used to process road images and detect road direction based on three directions: straight, left and right. IAfter several experiments, it was found that the slope of the line in the full photo could not represent the route of the road, therefore, the photos were divided into two, three and four parts and another effective parameter in the system. the seven layers were used to start the neural network training and each time the number of layers increased and the processing results were compared. The best answers were obtained for the number of layers above 70, which is a small difference between the photos divided into three and Four were observed and with a difference of 2%, the four-part photos gave us the best answer.