Adaptive Fusion of LiDAR and Camera Data for Enhanced Precision in 3D Object Detection for Autonomous Driving

Reliable 3D object detection is critical for autonomous driving, yet static fusion strategies often fail under occlusion, sparse point clouds, or varying illumination. This paper proposes the Adaptive Multi-Modal Fusion (AMMF) framework, which dynamically integrates LiDAR and camera features through three mechanisms: modality-specific gating to filter input relevance, multi-head attention for cross-modal interaction, and uncertainty modeling for confidence-aware weighting. Experiments on the KITTI benchmark demonstrate that AMMF achieves state-of-the-art performance for Cars and Cyclists and competitive results for Pedestrians, highlighting robustness in cluttered and low-visibility scenarios. Runtime analysis shows that the attention block dominates computation, limiting throughput to ~6.3 FPS; however, optimization strategies such as sparse attention or head sharing can substantially reduce latency. These findings indicate that AMMF not only improves detection accuracy but also establishes a scalable foundation for deploying multimodal perception in real-world autonomous driving systems.