A Rethinking of Real-Time Computer Vision-Based Lane Detection

Lane detection (LD) is one of the most important Advanced Driver Assistance Systems (ADASs) in modern vehicles. In this paper, a new solution for LD based solely on classical computer vision, dubbed HistWind, is presented. A great advantage of the proposed solution is real-time application, and no need for powerful hardware. HistWind is based on a combination of image preprocessing, histogram peak detection, and sliding window line detection and fitting. The parameters are adjusted using particle swarm optimization against an existing training data set and then compared with an existing and freely available deep learning-based solution on a testing set. The results show that HistWind achieves comparable results in terms of precision, recall, and F1 measures, but significantly lower execution time when operating on a CPU. This makes HistWind suitable for real-time implementation on limited hardware resources, which is often installed in modern vehicles. HistWind can process 25 frames per second when processing 1640x590 pixel frames and running on AMD Ryzen 5 1600 CPU with 24 GB of RAM.