Camera–LiDAR–Radar Multi-sensor Fusion for Autonomous Driving Perception: A Survey

Abstract

Camera, LiDAR, and radar are core sensing modalities of autonomous driving perception. Cameras provide dense appearance but inferred depth, LiDAR provides accurate 3D geometry but does not directly measure velocity and degrades in rain, fog, or dust, radar is relatively robust to adverse weather and measures Doppler velocity, yet returns sparse, noisy points. Fusion can therefore raise the accuracy and reliability of 3D detection, tracking, segmentation, and localization. A structured review of the 2019–2025 literature yields four findings. First, bird's-eye-view representation has emerged as a dominant fusion backbone because it removes the camera–LiDAR viewport mismatch that limited earlier proposal-level and ROI-pooling designs. Second, radar remains underused: the bottleneck often lies in its sparse and noisy point cloud rather than in the fusion operator itself. Third, spatial–temporal calibration and adverse-weather degradation set the practical performance ceiling more often than the fusion architecture itself. Fourth, the field is moving toward end-to-end and foundation-model paradigms. The survey organizes methods with a three-view taxonomy—processing stage, representation, and sensor pair—and maps these findings onto open problems and near-term research directions.