Automated Body Landmark Detection for Orthopedic Measurement Using RGB-D Cameras and Deep Learning

Abstract

Conventional tape-based anthropometry is examiner-dependent and prone to inter-rater vari-ability of 1.5-3.8 cm, motivating fully automated al-ternatives. Presented is an end-to-end deep-learning pipeline that localises eight clinically relevant body landmarks and computes the corresponding circumfer-ences from a single RGB-D frame captured with an Intel RealSense D415 camera. A PointTransformer re-gression network predicts 3-D landmark coordinates in under 80 ms on an NVIDIA RTX 3060 GPU; geodesic contours extracted from a fitted body mesh then yield circumference estimates within a mean absolute er-ror (MAE) of 0.8 cm on a hold-out set of 24 sub-jects (120 total). The system outperforms OpenPose-augmented depth baselines and reduces session time from several minutes to under five seconds. Ablation experiments validate the contributions of depth fusion, mesh regularisation, and the PointTransformer back-bone. An annotated evaluation dataset will be publicly released upon acceptance to support reproducibility.