3DLaneDT: 3D lane detection via reliable depth and temporal fusion for autonomous driving

Accurate 3D lane detection remains a critical challenge for autonomous vehicles, particularly in complex scenarios involving slopes and occlusions. While multimodal and temporal fusion paradigms have succeeded in general object detection, directly applying them to 3D lane detection is limited by the sparse and slender nature of lane markings. To address these limitations, we propose 3DLaneDT, a robust framework tailored for 3D lane detection. Distinct from existing depth-aware methods with standard convolutions, we propose an Adaptive Depth Network (ADN) incorporating Adaptive Area Convolution (AAConv). This design dynamically adjusts to the sparsity of lane depth signals, integrating reliable geometric priors into the Transformer attention mechanism. Furthermore, unlike attention-based temporal fusion which suffers from quadratic complexity, we introduce a State Space Model (SSM)-based Temporal Propagation Module (TPM). This module efficiently models the continuous geometric evolution of sparse lane queries. Experimental results on OpenLane and Apollo datasets demonstrate that 3DLaneDT surpasses state-of-the-art methods in both accuracy and efficiency. Code will be available after acceptance.