Presentation Information
[C-15-09]Filtering and Recovery of mmWave Radar Point Clouds Using Signal Features and Human-Body Structure
◎Yuka Tani1, Lee Chi-Hsuan2, Takuya Sakamoto1 (1. Kyoto Univ., 2. Quanta Computer Inc.)
Keywords:
mmWave radar,human sensing,human pose estimation,signal processing
In millimeter-wave radar-based human pose estimation, point cloud representations offer a lightweight data format highly compatible with 3D machine learning models. How- ever, due to their spatial sparsity, weakly reflecting parts such as limbs are often missing from observations. While deep learning can generate dense point clouds, it in- creases computational costs. We propose a method to gen- erate structurally informative point clouds with a limited number of points, using radar observation signals and initial point clouds generated by CLEAN. By combining reliability- based outlier removal and missing-part recovery based on signal features and human-body structure, our preprocess- ing approach aims to improve pose-estimation accuracy while reducing training and inference costs.