DFT-3DLaneNet: Dual-Frequency Domain Enhanced Transformer for 3D Lane Detection

In autonomous driving, 3D lane detection using a monocular camera presents significant challenges. In complex driving environments, spatial-domain image features may have certain limitations. To address these challenges, we propose a DFT-3DLaneNet model that uses a dual-frequency domain enhancement 3D unified deformable attention for monocular 3D lane detection. The model utilizes a learnable filter radius to extract dynamic high-frequency and low-frequency features, enhancing both the lanes and 3D space. On the one hand, we propose a dual-channel high-frequency feature enhancement module (DHF) to enhance lane features. On the other hand, we propose a cross-channel low-frequency attention module (CLA) to enhance 3D spatial perception. Low-frequency features not only alleviate the problem of imbalanced lane type distribution in the dataset but also improve the capture of 3D lane features and make the model more robust. Experimental results show that our method outperforms existing state-of-the-art approaches in terms of F-score.