OmniPC: A Generalizable Point Cloud Generation Pipeline for mmWave Radar

Point cloud data forms the foundation of mmWave sensing. However, existing point cloud generation pipelines utilize only a subset of radar dimensions - typically combining range with either Doppler velocity or super-resolution angle - thereby failing to fully exploit the capabilities of mmWave radar. The fundamental reason lies in the inherent conflict between Doppler velocity estimation and super-resolution angle estimation, which hinders their effective joint utilization. In this work, we introduce OmniPC, a unified point cloud generation pipeline that resolves this long-standing conflict and, for the first time, jointly exploits range, Doppler velocity, and super-resolution angle to produce higher-quality point clouds. The key idea is to redesign the processing pipeline using a sliding-window strategy, for which we provide both a comprehensive theoretical analysis and a practical deployment method. We implement OmniPC on three commercial mmWave radars operating at different frequency bands, demonstrating its strong generalizability across both frequency bands and radar devices. Extensive experiments show that OmniPC increases point-cloud density by 2.16 × compared to the research state-of-the-art and 3.99 × compared to the industrial state-of-the-art, while substantially enhancing the performance of downstream sensing applications.