Abstract: This paper presents a novel methodology for estimating vehicle lane deviation in driving and issuing timely alerts, in varying environmental lighting conditions such as night, day and rainy, while making Advanced Driver Assistance Systems (ADAS) improved and more accessible to all users. A Fully Convolutional Neural Network (FCNN), trained on dashboard camera video, is utilized for semantic segmentation to identify the drivable area under varying lighting and weather conditions. This is followed by lane deviation computation using a novel geometrical algorithm, ultimately producing lane deviation warning. Additionally, a U-Net model is integrated for generating lane lines in scenarios where visibility is impaired or lane markings are unclear. The proposed system has shown remarkably low inference times compared to existing models with similar accuracy in lane recognition. The departure prediction has shown exceptional specificity and negative predicted value. The proposed system, which uses a simplified algorithms and low-resolution inputs, is designed for low computational and resource deployments and can adapt to rainy, daytime, and nighttime environments. The proposed system has promising potential to strengthen road safety by providing drivers with a readily deployable ADAS, tailored for vehicles that do not have manufacturer-installed systems. Integrating real-time traffic and road condition information could enable proactive adjustments to driving behaviour, potentially avoiding accidents and paving the way for safer and more intelligent transportation systems.

Keywords: ADAS, U-Net, FCNN, Lane Departure Prediction, Rainy conditions, Night-time Driving

Pages: 51-59

WSEAS Transactions on