Improving Autonomous Vehicle Performance through Integration of an Image Deraining and a Deep Learning-Based Network for Lane Following

Abstract

—Lane-keeping is a vital component of autonomous driving that requires multiple artificial intelligence technologies and vision systems. However, maintaining a vehicle’s position within the lane is challenging when there is low visibility due to rain. In this research, a combination of image deraining and a deep learning-based network is proposed to improve the performance of the autonomous vehicle. First, a robust progressive Residual Network (ResNet) is used for rain removal. Second, a deep learning-based network architecture of the Convolutional Neural Networks (CNNs) is applied for lane-following on roads. To assess its accuracy and rain-removal capabilities, the network was evaluated on both synthetic and natural Rainy Datasets (RainSP), and its performance was compared to that of earlier research networks. Furthermore, the effectiveness of using both deraining and non-deraining networks in CNNs is evaluated by analyzing the predicted steering angle output. The experimental results show that the proposed model generates safe and accurate motion planning for lane-keeping in autonomous vehicles.