Density-Propagation Clustering and MLS Surface Reconstruction for 3D mmWave Radar Point Clouds

Millimeter-wave (mmWave) radar has become a key sensing technology for environmental perception and object detection under adverse weather and illumination conditions. However, radar point clouds are inherently sparse, noisy, and non-uniformly distributed, which makes reliable clustering and surface reconstruction challenging. This paper proposes a density-propagation clustering and surface reconstruction frame-work for mmWave radar point clouds that integrates K-nearest neighbor (KNN)-based local density estimation, KD-Tree acceleration, and Moving Least Squares (MLS) surface fitting. The method propagates cluster labels through local neighborhoods by jointly enforcing distance proximity and kNN-density similarity, enabling robust segmentation in heterogeneous radar data. The proposed framework was validated through both simulation and real-world experiments. In simulation, three synthetic soil mounds were generated to evaluate clustering accuracy and surface reconstruction quality. In real measurements, data were collected using a Texas Instruments AWR2243 four-cascade radar platform, where two soil mounds were reconstructed from top-down radar sensing. Experimental results show that the proposed algorithm achieves accurate and smooth 3D surface reconstruction, with sub-centimeter RMSE in simulation and centimeter-level accuracy in real radar data. Compared with DBSCAN, VDBSCAN, and REDBSCAN, the proposed method demonstrates improved robustness to density variations and measurement noise, while maintaining high computational efficiency, providing a practical solution for radar-based 3D modeling, terrain profiling, and environmental perception in autonomous and industrial applications.