RadarDiffusion: Camera Radar Fusion For 3D Object Detection

This thesis investigates whether diffusion models can enhance sparse radar point clouds for radar-only 3D object detection. Radar sensors provide long range, robustness to adverse weather, and direct velocity measurements, but their point clouds are sparse and noisy compared with LiDAR.

The work learns a mapping from radar observations to denser LiDAR-like bird's-eye-view representations using the nuScenes dataset. Radar and LiDAR point clouds are processed into BEV occupancy and height maps, then a conditional diffusion model reconstructs LiDAR BEV targets from radar inputs.

Generated representations are evaluated through a CenterPoint-based 3D detection pipeline. The results show that representation design is critical: preserving height information, BEV resolution, and preprocessing choices strongly affect downstream object detection performance.

The project combines multimodal perception, radar representation learning, diffusion-based generative modeling, and 3D object detection to explore how radar can become a more useful geometric sensing modality for autonomous driving.