Performance Comparison of Deep Learning Lane Detection Models for Autonomous Vehicles

Abstract

Recent advancements in the field of deep learning have significantly improved the driving capabilities of autonomous vehicles. This study focuses on the lane detection abilities of autonomous vehicles and examines the use of deep learning-based approaches in this context. The research compares the lane detection performance of various deep learning models, including U-Net, SCNN, ENet, and ENet-SAD, utilizing the TuSimple dataset. The models were evaluated using various quantitative metrics such as accuracy, precision, sensitivity, F1 score, and IoU. Extensive experiments have determined that the U-Net model exhibited the highest performance with an accuracy rate of 98.3%. The SCNN model, on the other hand, stood out in terms of precision, sensitivity, F1 score, and IoU metrics. In terms of inference time, the U-Net model was identified as the fastest lane detection model with a time of 20.12 ms. These results indicate that the U-Net model is particularly suitable for real-time systems requiring low computational power. Additionally, a qualitative assessment of lane detection success revealed that the SCNN and U-Net models more accurately detected pixels where lanes are present, whereas the ENet and ENet-SAD models were more prone to false-negative errors.