Multi-branch Coupled Dual Attention Network for Lane Detection

Lane detection is a key technology in autonomous driving and advanced driver assistance systems. Achieving high-precision detection while ensuring real-time performance remains challenging. This paper presents a novel lane detection approach integrating multi branch-coupled channel and spatial perception mechanisms to address the accuracy-speed trade-off challenge. Introducing a dynamic feature re-labeling module in high-level feature maps improves the model's adaptability to complex scenes while maintaining high inference speed. Our method is innovative and effective. It includes three key elements: Structural reparameterization strategy and inference stage optimization. This strategy enables complex multi-branch representation during training while ensuring efficient single-branch inference during deployment. Channel-dependent semantic enhancement mechanism. This models the dependencies between feature channels to adjust the activation patterns of each channel dynamically. Region attention mechanism based on spatial distribution. This generates spatial attention maps through lightweight convolution operations to enhance attention to local lane line regions. Our method improves the F1 score on the TuSimple dataset from 87.98% to 88.45% under the ResNet-34 architecture, with an accuracy rate of 96.50%. On the CULane dataset, the average F1 score reaches 73.57% while maintaining a real-time inference speed of 148.5 FPS. Ablation experiments definitively validate each component's effectiveness and synergistic effects, particularly in complex scenarios such as shadows, curves, and strong lighting conditions. This study presents a novel solution for efficient lane detection in autonomous driving systems.