DB3D-L: Depth-aware BEV Feature Transformation for Accurate 3D Lane Detection

TL;DR

This work tackles monocular 3D lane detection by enriching BEV feature construction with depth information. It introduces DB3D-L, a depth-aware BEV feature transformation framework consisting of D&F Net for joint depth and FV feature extraction, Fusion Net for depth-FV fusion into BEV via a cross-attention-like mechanism, and a BEV-based 3D lane detection head; losses combine depth and lane objectives. The method demonstrates competitive performance on ApolloSim and OpenLane, with notable gains in robustness and efficiency due to depth-guided BEV construction. Overall, the approach advances depth-aware BEV perception for reliable 3D lane detection in real-world driving scenarios.

Abstract

3D Lane detection plays an important role in autonomous driving. Recent advances primarily build Birds-Eye-View (BEV) feature from front-view (FV) images to perceive 3D information of Lane more effectively. However, constructing accurate BEV information from FV image is limited due to the lacking of depth information, causing previous works often rely heavily on the assumption of a flat ground plane. Leveraging monocular depth estimation to assist in constructing BEV features is less constrained, but existing methods struggle to effectively integrate the two tasks. To address the above issue, in this paper, an accurate 3D lane detection method based on depth-aware BEV feature transtormation is proposed. In detail, an effective feature extraction module is designed, in which a Depth Net is integrated to obtain the vital depth information for 3D perception, thereby simplifying the complexity of view transformation. Subquently a feature reduce module is proposed to reduce height dimension of FV features and depth features, thereby enables effective fusion of crucial FV features and depth features. Then a fusion module is designed to build BEV feature from prime FV feature and depth information. The proposed method performs comparably with state-of-the-art methods on both synthetic Apollo, realistic OpenLane datasets.

Figures (6)

• Figure 1: The pipeline of previous BEV lane detection (upper) and our proposed DB3D-L (lower). Previous methods mainly utilize camera parameters and inverse perspective mapping (IPM) to transform the features into BEV space. Our methods build BEV by depth feature and FV features fusion with attention mechanism. Note that only standard operators are adopted in DB3D-L.
• Figure 2: FV features (a) vs BEV features (b). We find that BEV features and FV features have little difference essentially, they are both N channels 2-dimension value map. So we build BEV features just as a features map of CNN outputs. The FV features and BEV features has the common dimension in width, which is critical for the transformation from FV to BEV.
• Figure 3: The overall architecture of our proposed method DB3D-L. There are there main modue, D&F Net, Fusion Net, and Lane Head. In D&F Net, SE Layer is used to realize space cross attention, in Fusion Net, PFE and DAT is used to reduce the height dimension of the two types features.
• Figure 4: Image space representation, pseudo point clouds and BEV space representation. There are key elements overlap in height dimension in image space but no overlap in others. $\dagger$ represents key element. 1 represents Perspective. 2 represents IPM. 3 represents orthographic projection from top.
• Figure 5: The process of fuse BEV frature from prime FV feature and prime depth feature.