Context-guided feature enhancement for robust monocular 3D lane detection

Robust 3D lane detection from monocular images is essential for autonomous driving, as it provides accurate spatial lane geometry without relying on depth sensors or bird’s-eye-view transformations. However, this task remains highly challenging in real-world conditions due to factors such as occlusions, varying road curvature, perspective distortion, and complex lighting environments. To address these issues, we propose a feature enhancement approach tailored for geometry-aware and illumination-resilient 3D lane detection. Our method introduces three key components: (1) an Adaptive Scale-aware Fusion Module (ASFM), which dynamically reweights and fuses multi-scale backbone features via channel attention and dilation-adaptive convolutions to better integrate scene-dependent semantics; (2) a Context-Aware Feature Enhancement (CAFE) module, which incorporates frequency-domain priors and brightness-guided modulation to refine high-level features with improved structural consistency and photometric adaptability; and (3) a Spatially-Enhanced Instance Decoding (SEID) module, which strengthens instance-level lane prediction by embedding directional feature modeling and applying dual-stage attention refinement to enhance spatial coherence and suppress irrelevant activations. These modules are lightweight and complementary, forming a robust enhancement pipeline that significantly improves the quality of semantic and geometric representations. Extensive experiments on the OpenLane and Apollo 3D Synthetic datasets validate the effectiveness of our approach, showing consistent performance gains over competitive baselines, especially in complex scenarios involving illumination variation, partial occlusion, and lane curvature.