Human Fall Detection

Pose Estimation

LiteRT (TensorFlow Lite)

HRNet

This unit implements a fully heuristic, rule-based fall-detection algorithm that operates on a single human pose extracted from a static frame. It relies on skeletal keypoints generated by upstream pose-estimation models (e.g., HRNet) and evaluates the body geometry to determine whether a person is in a fallen state. The detection pipeline includes two logic paths: (1) a high-priority horizontal-pose check based on the bounding-box aspect ratio, and (2) a majority-voting system that combines multiple geometric indicators such as vertical compression, low-head/low-hip positioning, shoulder-tilt angle, and pose instability. These rules are designed to maximize sensitivity to real falls while reducing false positives in ambiguous poses. The system includes normalization of keypoints, validation of missing/low-confidence landmarks, and configurable thresholds to adapt the heuristic behavior. As a rule-based model, it is lightweight, requires no ML inference, and is highly suitable for low-power edge devices. Its effectiveness depends strongly on the quality of upstream pose detection and does not incorporate temporal motion analysis.