Comparative Transfer Learning Models for End-to-End Self-Driving Car

Abstract

Self-driving automobiles are prominent in science and technology, which affect social and economic development. Deep learning (DL) is the most common area of study in artificial intelligence (AI). In recent years, deep learning-based solutions have been presented in the field of self-driving cars and have achieved outstanding results. Different studies investigated a variety of significant technologies for autonomous vehicles, including car navigation systems, path planning, environmental perception, as well as car control. End-to-end learning control directly converts sensory data into control commands in autonomous driving. This research aims to identify the most accurate pre-trained Deep Neural Network (DNN) for predicting the steering angle of a self-driving vehicle that is suitable to be applied to embedded automotive technologies with limited performance. Three well-known pre-trained models were compared in this study: AlexNet, ResNet18, and DenseNet121. Transfer learning was utilized by modifying the final layer of pre-trained models in order to predict the steering angle of the vehicle. Experiments were conducted on the dataset collected from two different tracks. According to the study's findings, ResNet18 and DenseNet121 have the lowest error percentage for steering angle values. Furthermore, the performance of the modified models was evaluated on predetermined tracks. ResNet18 outperformed DenseNet121 in terms of accuracy, with less deviation from the center of the track, while DenseNet121 demonstrated greater adaptability across multiple tracks, resulting in better performance consistency.