Drone + Wearable Sensor Fusion Enables 85.7% Accurate Battlefield Triage Without Direct Casualty Access
ATRACT is a human-in-the-loop decision support system that fuses drone-captured video with wearable physiological sensors to perform remote casualty triage in contested battlefield environments. The system uses multi-modal learning — combining behavioral cues (pose, posture) from drone video with heart rate, breathing rate, and movement data from body-worn sensors — to assist frontline medics when direct casualty access is delayed or dangerous. A conditional variational autoencoder (CVAE) is used to bridge the data realism gap for injured action sequences, enabling training on a drone-captured dataset. The pipeline achieves 85.7% action classification accuracy using a lightweight CNN visual encoder that remains competitive with heavier pre-trained video backbones.