Realization of the Autonomous Driving System on the Experimental Vehicle

Authors : Namig ALİYEV, Mehmet Turan GÜZEL, Oğuzhan SEZER

Pages : 48-56

Doi:10.21541/apjess.1060763

View : 12 | Download : 6

Publication Date : 2022-01-01

Article Type : Research Paper

Abstract :Running control software on limited computing resources is considered one of the toughest problems. In this study, an autonomous driving software has been developed that can safely complete the map by tracking the lanes and avoiding obstacles on a robot vehicle with limited hardware components. The data was simplified with the image processing technique and the neural network was trained. Overfitting was prevented by hyperparameter tuning and synthetic data augmentation. In order to avoid obstacles, optical flow was calculated by detecting corners every 4 seconds and was used to find the focus of expansion of the vehicle. Time-to-collision was found with the FOE and the distance between the previous position and the current position of the detected point. Optimization was made by averaging the values of close points. The balance mechanism was created according to the TTC difference calculated on the right and left parts of the vehicle.
Keywords : Convolutional Neural Network, Overfitting, Hyperparameter Tuning, Data augmentation, Lane tracking, Optical Flow, Focus of Expansion, Time to Collision