<p>To address the difficulty of traditional convolutional neural networks (CNNs) in detecting occluded lanes and curves caused by their inability to fully integrate continuous and slender lane structures, this paper proposes an improved lane detection network integrating deformable dilated convolution and feature pyramid. First, deformable and dilated convolutions are introduced into Layer4 of the backbone network. This allows for the adaptive adjustment of the receptive field to accurately capture the spatial structure of slender lanes. Second, an FPN multi-feature fusion module enables top-down and bottom-up cross-layer fusion of low-level high-resolution and high-level semantic features, significantly enhancing reasoning capabilities for occluded and distant incomplete lanes. Furthermore, an optimized combined loss function is constructed by jointly utilizing Focal Loss and clustering-based discriminative loss. This effectively alleviates the extreme imbalance between lane pixels and background pixels, further improving the classification accuracy of lane lines. The verification and analysis on the Tusimple dataset, the proposed method achieves 98.1% accuracy. Finally, benefiting from the lightweight ResNet-50 backbone and a local optimization strategy that introducing the improved module solely into Layer4, the proposed design effectively limits computational overhead and inference latency while maintaining high detection precision. This demonstrates substantial potential for lightweight deployment on in-vehicle terminals.</p>