Evaluation of Convolutional Neural Networks for Predicting Steering Angles in Autonomous Driving Systems Using the Udacity Simulator

Abstract

This research explores the use of convolutional neural networks (CNNs) to develop autonomous driving systems, focusing on predicting steering angles for autonomous vehicles. A custom dataset was created using the Udacity simulator, which provides a high-fidelity simulation environment with multiple lanes and three-angle cameras to collect driving data. This data encompasses, but is not limited to, vehicle speed, steering angle, and gear position. To achieve this objective, two CNN models were utilized: the simple Comma.ai model and the more complex NVIDIA model. These models were utilized to compare performance and examine their ability to predict steering angles accurately. The experimental framework entailed the training of models on data using mean loss evaluation (MSE) and the subsequent validation of their performance on independent validation data. The findings of the study demonstrated that both models exhibited the capacity to differentiate driving patterns from visual data. However, the NVIDIA model exhibited superior performance in complex environments, a feat attributable to its advanced architecture and its ability to extract more accurate features. Conversely, the Comma model demonstrated superior performance in less complex environments, making it a suitable option for simpler systems.