Lane Detection on Structured Roads using Ternaus-Net Architecture

Lane detection is considered one of the essential elements that an autonomous driving system must provide accurate navigation over roads to ensure the safety of vehicles. Current lane detection techniques are still not good enough, and they cannot recognize lanes in structured and semi-structured environments, especially when applied in real-world driving environments. This paper introduces a U-Net-based architecture named TernausNet, which adopts an encoder-decoder based on the encoder-decoder version of VGG16, particularly fine-tuned for accuracy in segmentation tasks. This architecture is part of a deep-learning approach toward lane detection. TuSimple is a quite popular benchmark for lane detection, which covers all sorts of road conditions from around the world. The model is trained and assessed using this dataset. It can demonstrate that even in challenging circumstances, like faded lane lines, shadows, and occlusions, the results are dependable and accurate in identifying lane markings.