Reproduce “High Resolution Point Clouds from mmWave Radar” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Akarsh Prabhakara, Tao Jin, Arnav Das, Gantavya Bhatt, Lilly Kumari, Elahe Soltanaghaei, Jeff Bilmes, Swarun Kumar and 1 more
A deep learning system, RadarHD, generates high-resolution, lidar-like point clouds from low-resolution single-chip mmWave radar inputs, maintaining performance in dense smoke where lidar fails and achieving a 3.5x improvement in modified-Hausdorff error over traditional radar processing. This enables robust environmental perception in conditions where optical sensors are ineffective.
Key insights
- Exploiting a low-threshold representation of spatially Fourier transformed raw I/Q data allows the deep learning model to discern true objects from complex radar artifacts.
- Stacking multiple past radar frames provides temporal context, enabling the network to overcome specularity, maintain object persistence, and infer true objects across viewpoints.
- An asymmetric U-Net architecture is effective for super-resolution by learning to denoise and extract accurate semantic representations while capturing a global understanding of radar images to mitigate widespread artifacts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.