FGP-BEVFusion for Lightweight Camera–LiDAR 3D Object Detection

Abstract

Reliable 3D object detection is essential for autonomous-driving perception in complex traffic environments. Camera–LiDAR fusion combines dense visual semantics with precise point-cloud geometry to improve detection robustness. BEVFusion is a competitive bird’s-eye-view (BEV) fusion framework, but its high-capacity image backbone introduces considerable computational overhead. Directly substituting a lightweight backbone often degrades visual representations and reduces detection accuracy. To address this accuracy–efficiency dilemma, we construct Feature Generation and Propagation BEVFusion (FGP-BEVFusion), a lightweight camera–LiDAR 3D object detector. Multi-scale Efficient Channel-aware Ghost convolution (MEC-Ghost) strengthens the lightweight backbone through identity-initialized multi-scale depthwise mappings and efficient channel recalibration. Bidirectional Path-Aggregation Semantic–Spatial Feature Pyramid Network (BPA-SSFPN) aggregates C2 to C5 features through top-down, bottom-up, and cross-level paths. This design addresses weakened spatial localization and insufficient high-level semantic propagation in lightweight architectures. On nuScenes validation, the framework attains a mean average precision (mAP) of 0.679 and a nuScenes detection score (NDS) of 0.710. These scores surpass the reproduced Swin-T baseline by 0.003 and 0.002, respectively, under identical experimental settings. It also reduces total parameters by 32.12 percent and profiler-counted floating-point operations by 15.30 percent, while raising single-frame throughput by approximately 4.06 percent on an RTX 4090. These results provide a favorable balance between detection accuracy and computational cost for lightweight multimodal 3D object detection.