Deep learning–based estimation of entrapment status of survivors under debris from 3D measurements for disaster medical response

Severe earthquakes, landslides, and tsunamis can cause structural collapse, trapping survivors beneath debris and limiting direct observation of the entire body. In confined spaces, voice calls, still images, and videos may be inadequate for conveying a survivor’s posture, contact with debris, and surrounding spatial conditions. We therefore developed a deep learning–based pipeline that takes a textured 3D scene mesh as input and integrates multiview, disaster-adapted human pose and shape estimation, dual-label RGB-D segmentation, scene-constrained SMPL fitting, and voxel-based geometric reasoning. The pipeline reconstructs a full-body survivor mesh, identifies candidate debris–body contact regions, and estimates the mass of load-contributing debris using a simplified geometric criterion. We evaluated the pipeline in 23 controlled entrapment cases involving clothed mannequins and representative wood and concrete blocks. Across these cases, the complete pipeline achieved a contact-region intersection-over-union of 0.391, an F1-score of 0.538, and a relative debris-mass error of 26.1%. Although these results do not establish the method’s suitability for clinical diagnosis, crush-syndrome severity assessment, or autonomous medical or rescue decision-making, they demonstrate its technical feasibility for generating preliminary situational information from 3D scene measurements.