Unsupervised Radar Point Cloud Enhancement via Arbitrary LiDAR Guided Diffusion Prior

TL;DR

This work tackles the limited angular resolution of radar by eliminating the need for paired LiDAR-radar training data. It introduces an unsupervised framework that uses a LiDAR-domain diffusion prior as a Bayesian regularizer in solving the radar angle estimation inverse problem, enabling posterior sampling of enhanced radar points. A LiDAR-trained latent diffusion model provides the prior, while the radar imaging forward function acts as the likelihood, yielding LiDAR-consistent yet radar-faithful outputs. Experiments on the RADIal and K-Radar datasets demonstrate competitive performance with supervised methods and strong cross-domain generalization, though the method incurs significant inference cost and requires careful multi-modal alignment in practice.

Abstract

In industrial automation, radar is a critical sensor in machine perception. However, the angular resolution of radar is inherently limited by the Rayleigh criterion, which depends on both the radar's operating wavelength and the effective aperture of its antenna array.To overcome these hardware-imposed limitations, recent neural network-based methods have leveraged high-resolution LiDAR data, paired with radar measurements, during training to enhance radar point cloud resolution. While effective, these approaches require extensive paired datasets, which are costly to acquire and prone to calibration error. These challenges motivate the need for methods that can improve radar resolution without relying on paired high-resolution ground-truth data. Here, we introduce an unsupervised radar points enhancement algorithm that employs an arbitrary LiDAR-guided diffusion model as a prior without the need for paired training data. Specifically, our approach formulates radar angle estimation recovery as an inverse problem and incorporates prior knowledge through a diffusion model with arbitrary LiDAR domain knowledge. Experimental results demonstrate that our method attains high fidelity and low noise performance compared to traditional regularization techniques. Additionally, compared to paired training methods, it not only achieves comparable performance but also offers improved generalization capability. To our knowledge, this is the first approach that enhances radar points output by integrating prior knowledge via a diffusion model rather than relying on paired training data. Our code is available at https://github.com/yyxr75/RadarINV.

Figures (15)

• Figure 1: Schematic diagram of the radar angle measurement principle. The five-pointed star symbolizes objects detected in space. As the object-radar distance greatly exceeds the inter-antenna distance, the incident electromagnetic field is treated as a plane wave, arriving at each antenna at the same angle.
• Figure 2: Diffusion prior for unsupervised radar point cloud enhancement. Let $z_t$ be the latent at step $t$, with $\mathcal{E}$, $\mathcal{D}$ as the autoencoder. The forward and reverse diffusion are modeled by $q(\boldsymbol{z}_t \mid \boldsymbol{z}_{t-1})$ and $p_{\theta}(\boldsymbol{z}_{t-1} \mid \boldsymbol{z}_t)$. ${A}(\cdot)$ is the radar measurement system forward function; - denotes element-wise subtraction. (a) A diffusion model is trained on LiDAR latents to learn a prior. (b) Enhancement samples from the posterior combining LiDAR prior with radar input. (c) Each reverse step adds an $L_2$ gradient to enforce radar fidelity. (d) This gradient is computed between ${A}(\bar{\boldsymbol{x}}_0^{(k)})$ and radar data $\boldsymbol{Y}$ w.r.t. $\hat{\boldsymbol{z}}_{t-1}$. (b) to (d) constitute the inference process, which only requires $p_{\theta}(\boldsymbol{z}_{t-1} \mid \boldsymbol{z}_t)$ estimated from a trained model in (a), no more training is needed. The details of process (a), (b) are described in section \ref{['Diffusion_Forward_Process']} and \ref{['Diffusion_Reverse_Process']}, (c) and (d) are further explained in section \ref{['Using_Diffusion_Model_as_a_Prior_for_Enhancement']}.
• Figure 3: Qualitative comparison of the RADIal dataset across different methods. Different randomly selected frames from three scenarios are displayed in rows, while point enhancement results from different methods are shown in columns. In each sub-figure, blue points represent the LiDAR ground truth, while red points indicate the enhanced point cloud outputs. More results please refer to Figure \ref{['fig:appendix_vis_radial']} in Appendix \ref{['app_g']}.
• Figure 4: Cross-dataset enhancement on K-Radar. Rows show radar inputs; columns show outputs from different methods. Blue and red points denote LiDAR (GT) and generated points respectively. Other methods struggle with noise and domain shifts, while our approach yields cleaner, robust results without requiring paired training data. More results please refer to Figure \ref{['fig:appendix_vis_kradar']} in Appendix \ref{['app_g']}.
• Figure 5: Visualization results of cross-scenario validation on RADIal dataset. Three rows demonstrate three scenarios provided by RADIal dataset. In each scenario, two frames are selected. Results of all as prior and countryside as prior are provided. More results please refer to Figure \ref{['fig:appendix_vis_cross_scenario']} in Appendix \ref{['app_g']}.