Hybrid MambaVision and Transformer-Based Architecture for 3D Lane Detection

Lane detection is an essential task in the field of computer vision and autonomous driving. This involves identifying and locating road markings on the road surface. This capability not only helps drivers keep the vehicle in the correct lane, but also provides critical data for advanced driver assistance systems and autonomous vehicles. Traditional lane detection models work mainly on the 2D image plane and achieve remarkable results. However, these models often assume a flat-world scenario, which does not correspond to real-world conditions, where roads have elevation variations and road markings may be curved. Our approach solves this challenge by focusing on 3D lane detection without relying on the inverse perspective mapping technique. Instead, we introduce a new framework using the MambaVision-S-1K backbone, which combines Mamba-based processing with Transformer capabilities to capture both local detail and global contexts from monocular images. This hybrid approach allows accurate modeling of lane geometry in three dimensions, even in the presence of elevation variations. By replacing the traditional convolutional neural network backbone with MambaVision, our proposed model significantly improves the capability of 3D lane detection systems. Our method achieved state-of-the-art performance on the ONCE-3DLanes dataset, thus demonstrating its superiority in accurately capturing lane curvature and elevation variations. These results highlight the potential of integrating advanced backbones based on Vision Transformers in the field of autonomous driving for more robust and reliable lane detection. The code will be available online.