Feature extraction and clustering of lower-body shapes in older women using unsupervised contrastive learning on 3D point clouds

Significant variations exist in the lower-body shapes of older women. Traditional anthropometric analysis methods cannot capture their complex three-dimensional (3D) morphology. This study proposes an unsupervised learning framework that extracts lower-body shape features from raw 3D point-cloud data for shape-based clustering analysis. 3D body scans of 245 women aged 60–80 years were collected, and a proportional thresholding method combined with density-based spatial clustering of applications with noise (DBSCAN) clustering was employed to segment anatomical regions and construct a standardized 3D morphological dataset. PointNet, a deep learning network designed for unstructured point clouds, was then used to extract a 1024-dimensional global shape feature vector for each individual. To improve feature discriminability and robustness, the InfoNCE contrastive loss function was introduced during training, encouraging closer representations of similar samples while separating dissimilar samples. Random geometric augmentations, including rotation, scaling, and noise injection, were applied to enhance model generalizability. On the basis of the extracted features, k -means clustering grouped the samples into three clusters interpreted as typical lower-body shape types according to their morphological characteristics: (1) convex waist and hips (30.61%), (2) protruding abdomen with rounded hips (23.27%), and (3) convex waist with flat hips (46.12%). Clustering performance was validated using a silhouette score of 0.394, a Calinski–Harabasz index of 278.904, and a Davies–Bouldin index of 0.837. This method requires no manual measurements or predefined landmarks and provides a data-driven basis for trouser pattern development, personalized sizing recommendations, and virtual try-on systems. It also offers a methodological reference for 3D-shape analysis of other populations and body regions.