Curvature Entropy Enhanced Principal Component Analysis for 3D LiDAR Human Head Segmentation

Abstrak

Accurate segmentation of the human head and torso from Three-Dimensional (3D) LiDAR point clouds is essential for applications in human–robot interaction, autonomous navigation, and contactless rehabilitation monitoring. Traditional unsupervised approaches based solely on Principal Component Analysis (PCA) often suffer from instability under pose variation and sensor noise, as the geometric width of the head and shoulders can overlap. This study proposes a novel Curvature Entropy-Enhanced Principal Component Analysis (PCA+CE) framework that combines global geometric width descriptors with local surface CE to achieve robust, unsupervised head–body segmentation. Specifically, a CE term was computed from surface-normal directional randomness and fused adaptively with PCA width through a lightweight weighting-and-thresholding scheme to estimate the head–torso boundary automatically, without labeled data or learning. Experiments using human subsets of the KITTI Raw 3D LiDAR dataset demonstrate that PCA+CE consistently improves segmentation accuracy over the pure PCA baseline, achieving higher precision, recall, and F1-score. The method yields a 0.36 % increase in overall F1-score from 84.1% to 84.4% while maintaining computational efficiency and full interpretability. These results highlight CE as an effective local stabilizer for lightweight geometric segmentation in sparse LiDAR environments.

Identifier

• DOI: https://doi.org/10.48084/etasr.17551
• OpenAlex: https://openalex.org/W7150001935