Self-Supervised Diffusion-Based Scene Flow Estimation and Motion Segmentation With 4D Radar

Scene flow estimation (SFE) and motion segmentation (MOS) using 4D radar are emerging yet challenging tasks in robotics and autonomous driving applications. Existing LiDAR- or RGB-D-based point cloud processing methods often deliver suboptimal performance on radar data due to radar signals' highly sparse, noisy, and artifact-prone nature. Furthermore, for radar-based SFE and MOS, the lack of annotated datasets further aggravates these challenges. To address these issues, we propose a novel self-supervised framework that exploits denoising diffusion models to effectively handle radar noise inputs and predict point-wise scene flow and motion status simultaneously. To extract key features from the raw input, we design a transformer-based feature encoder tailored to address the sparsity of 4D radar data. Additionally, we generate self-supervised segmentation signals by exploiting the discrepancy between robust rigid ego-motion estimates and scene flow predictions, thereby eliminating the need for manual annotations. Experimental evaluations on the View-of-Delft (VoD) dataset and TJ4DRadSet demonstrate that our method achieves state-of-the-art performance for both radar-based SFE and MOS. The code and pre-trained weights of our method will be released at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/nubot-nudt/RadarSFEMOS</uri>.