BEV-ConvFusion: An Efficient 2D Fusion Framework for Real-Time Autonomous Perception

Engineering notes

By eliminating 3D operations, BEV-ConvFusion achieves substantial computational savings while maintaining high accuracy. Extensive experiments demonstrate that our method achieves competitive detection accuracy, significantly higher frame rates, and superior robustness compared to unimodal baselines, highlighting the effectiveness of 2D-domain fusion.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Autonomous driving requires perception systems that balance accuracy with computational efficiency for safe and reliable real-time operation. While camera–LiDAR fusion has emerged as a powerful solution, most existing methods rely on computationally expensive 3D backbones, limiting deployment on resource-constrained vehicle hardware. We propose BEV-ConvFusion, a novel 2D-domain fusion framework that overcomes this limitation. Our approach first encodes sparse LiDAR point clouds into dense multi-channel Bird’s-Eye View (BEV) representations and extracts semantically rich features from RGB images using a 2D CNN backbone. At the core of our design is the Synergistic Cross-Attention Module (SynCAM), which refines features through three sequential stages: spatial gating, bidirectional semantic cross-attention, and feature refinement, enabling reciprocal enhancement between modalities before fusion. By eliminating 3D operations, BEV-ConvFusion achieves substantial computational savings while maintaining high accuracy. Extensive experiments demonstrate that our method achieves competitive detection accuracy, significantly higher frame rates, and superior robustness compared to unimodal baselines, highlighting the effectiveness of 2D-domain fusion.

Links and sources

• Official / arXiv page

Comments