SUPER: A Novel Lane Detection System

AI-based lane detection algorithms were actively studied over the last few years. Many have demonstrated superior performance compared with traditional feature-based methods. However, most methods remain riddled assumptions and limitations, still not good enough for safe reliable driving in real world. In this paper, we propose a novel system, called Scene Understanding Physics-Enhanced Real-time (SUPER) algorithm. The proposed method consists of two main modules: 1) hierarchical semantic segmentation network as scene feature extractor 2) physics enhanced multi-lane parameter optimization module inference. We train system using heterogeneous data from Cityscapes, Vistas Apollo, evaluate on four completely separate datasets (that never seen before), including Tusimple, Caltech, URBAN KITTI-ROAD, Mcity-3000. approach performs same or better than models already trained dataset, well even that was on. Real-world vehicle tests also conducted. Preliminary test results show promising real-time lane-detection Mobileye.