Two-stage anthropometric landmark detection framework on human point cloud:-invariant and static-specific

초록

The rapid advancement of wearable and personalized devices has increased the demand for analyzing human shapes and poses from point cloud data. Traditional three-dimensional (3D) anthropometric landmark detection methods rely solely on static pose scans, limiting their effectiveness in dynamic postures. To address this limitation, we propose a two-stage 3D anthropometric landmark detection framework to enhance landmark detection accuracy for both dynamic and static postures, using artificial intelligence (AI) with deep learning. In the first stage, the model performs shape and pose invariant landmark detection from a single scan using a parametric human template model. Then, in the second stage, it performs refined landmark detection using point-specific segments for the input point cloud. To achieve these capabilities, we present a novel feature extractor specialized for anthropometric landmark detection, which extracts landmark-specific details. Our method outperforms existing approaches on a comprehensive, large-scale practical dataset collected in South Korea, consistently maintaining high accuracy across various dynamic poses.

키워드