A Depth-Centric Review of LiDAR–Camera Fusion: From 2D Sensors to High-End Perception for Autonomous Systems

Existing surveys on LiDAR–camera fusion share a blind spot: they assume multi-beam 3D LiDAR as the default sensor, overlooking the wide range of real-world configurations, from a single-plane 2D LiDAR on an indoor robot to a multi-sensor suite on a self-driving car. This review argues that the optimal fusion strategy is determined not by the highest-end architecture, but by the depth information actually available. We introduce the concept of a depth continuum and a three-axis taxonomy that maps any LiDAR–camera configuration (defined by LiDAR beam count and camera depth capability) to its appropriate fusion strategy: projection-based, depth completion, Bird’s Eye View (BEV), attention-based, or late fusion. Through a structured analysis of the 2021–2026 literature across autonomous driving, industrial AGVs, agricultural robotics, and indoor mobile robots, we derive five key findings: (i) the fusion strategy should track the available depth, with projection and depth completion for sparse LiDAR, BEV for moderate density, and attention-based fusion for dense configurations; (ii) depth completion is the bridging mechanism that supplies low-cost setups with dense depth, but only when trained at the target LiDAR density; (iii) BEV fusion is validated in the literature only from about 32 beams upward, a practical lower bound rather than a controlled threshold; (iv) no standardized benchmark exists for 2D LiDAR–camera fusion, creating a gap between academic research and practical robotics; and (v) late fusion, often dismissed as simplistic, keeps operating when one sensor is lost, a robustness that follows from its architecture but has not yet been validated on large-scale benchmarks. This review aims to serve both as a structured reference for researchers and as a practical configuration-to-strategy mapping for engineers designing perception systems across the full spectrum of autonomous applications.