A Comprehensive Survey of Deep Learning and Multimodal Fusion for 3D Object Detection in Autonomous Driving

Autonomous driving technology has seen remarkable evolution, propelled by advances in deep learning and multi-sensor fusion. This review traces the development of autonomous driving systems from early rule-based models in the 1980s to contemporary end-to-end frameworks built on large-scale foundation models. A key focus is the shift from traditional modular pipelines to unified architectures that incorporate transformers and bird's-eye view (BEV) representations, as seen in methods such as BEVFormer, TransFusion, and UniTR. The review also discusses widely used benchmark datasets—including KITTI, the Waymo Open Dataset, and nuScenes—as well as standard evaluation metrics such as 3D mAP, NDS, and system latency. Experimental comparisons show that multimodal fusion approaches, particularly BEVFusion and UniTR, deliver state-of-the-art accuracy while maintaining reasonable computational efficiency. Nonetheless, challenges remain, especially in detecting small objects and achieving real-time performance. These limitations highlight the need for lightweight, adaptable models. Overall, this review synthesizes recent progress, outlines unresolved research questions, and offers perspectives on future directions toward scalable, safe, and intelligent autonomous driving systems.