Deep-estimate excluded BEV detection with Two stage LiDAR and Camera Fusion

Bird’s-Eye View (BEV) representation is a core paradigm for 3D perception in autonomous driving. Vision-based BEV methods often rely on depth estimation, which introduces noise and computational overhead, while LiDAR provides accurate geometry but lacks semantic richness and is limited by data sparsity. We propose a novel hybrid framework that directly fuses image and LiDAR features without relying on explicit depth estimation. A lightweight cross-attention Transformer maps image features to BEV space, and a two-stage fusion module progressively integrates multi-modal features. Our method outperforms depth-based and early-fusion baselines, achieving 72.2% mAP on the nuScenes benchmark. The proposed architecture is efficient, robust, and well-suited for real-time deployment in autonomous systems.