DTANet: Dynamic Topology-Aware Network for Lane Detection in Complex Scenarios

Robust and accurate lane detection remains a critical challenge in autonomous driving, particularly under complex road scenarios such as occlusions, intersections , and curved lanes. Existing methods often struggle to preserve the topological structure of lanes due to limited geometric modeling capacity. To address this, we propose the Dynamic Topology-Aware Network (DTANet), which explicitly models and dynamically adapts the topology of lane during learning: it maintains the continuity of the lane under occlusions via adaptive contextual aggregation and accurately models curved lanes using differential geometric constraints. Specifically, DTANet has three core components: a Dynamic Topology Encoder (DTE) transforms spatial features into a graph-based representation , capturing local-global topological dependencies via multi-scale graph convolutions and dynamic attention; a Topology Distillation Enhancer (TDE) combines an adaptive gated mechanism with a geometric constraint—this constraint aligns curvature and tangent direction fields across feature levels to avoid semantic enhancement eroding lane geometric integrity, and the gated mechanism adjusts distillation pathways by scene complexity to reinforce continuity in degraded regions; a Topology Consistency Loss (TCL) supervises predicted curves via differential geometric constraints and penalizes curvature and orientation discrepancies against ground truth to provide global topological supervision. Experimental results on challenging lane detection benchmarks indicate that our method demonstrates promising performance compared against state-of-the-art models.