Diffusion-Based Reconstruction of 3D Occupancy Maps from 4D Radar Tensors

4D radar sensing provides rich measurements across doppler, range, azimuth, and elevation dimensions, offering strong resilience in adverse environmental conditions. However, reconstructing accurate 3D occupancy maps from 4D radar tensors (4DRT) remains challenging due to low spatial resolution, nonuniform sampling, and artifacts caused by sidelobes and multipath reflections. This paper proposes a diffusion-based framework for 3D occupancy reconstruction that directly leverages the encoded structure of 4DRT data. The pipeline consists of three components: a feature-preserving dimensionality reduction module that produces doppler-aware sparse descriptors from raw 4DRT; a hierarchical pillar-based representation that encodes vertical geometry using normalized height segments within a compact and structured spatial format; and a conditional diffusion model that iteratively denoises latent occupancy predictions. To encode the sparse and irregular 4DRT inputs, we design a hybrid condition encoder that combines convolutional layers with self-attention to extract both local and global contextual features. These are injected into a U-Net-based denoising network via cross-attention to generate dense and spatially consistent 3D occupancy volumes. Extensive experiments on the Coloradar benchmark and real-world driving data demonstrate that the proposed method consistently outperforms the state-of-the-art baseline, improving IoU from at most 3.3% to over 30% across diverse scenes, while reducing the Chamfer Distance by more than 4× and maintaining a single-frame inference latency of approximately 50 ms. The implementation is available at https://github.com/MoYuGit/Diffusion_radar_reconstruction.