Steering Angle Prediction for Autonomous Vehicles Using Deep Transfer Learning

Abstract

Self-driving cars are anticipated to revolutionize future transportation due to their reliability, safety, and continuous learning capabilities. Researchers are actively engaged in developing autonomous driving systems, employing techniques like behavioral cloning and reinforcement learning. This study introduces a distinctive perspective by employing an end-to-end approach, using camera inputs to predict steering angles based on a model learned from human driving expertise. The model demonstrates rapid training and achieves over 90.1% accuracy in Mean Percentage of Prediction (MPP). In this context, the study aims to replicate driver behavior by applying transfer learning from a pre-trained VGG19 model with various activation functions. The proposed model is trained to analyze road images as input, predicting optimal steering adjustments. Evaluation encompasses a dataset from the ROS2 simulation environment, comparing results with several Convolutional Neural Networks (CNNs) models including Nvidia's, MobileNet-V2, ResNet50, VGG16, and VGG19. The impact of activation functions like Exponential Linear Unit (ELU), Rectified Linear Unit (ReLU), and Leaky ReLU on the transfer learning model is also explored. This research contributes to advancing autonomous driving systems by addressing real-world driving complexities and facilitating their integration into everyday transportation. The novel approach of utilizing transfer learning and comprehensive evaluation underscores its significance in optimizing self-driving technology.