wtf is real-time-fall-detection-system?

This is a fall detection system that runs entirely on a low-power edge device called the NVIDIA Jetson Nano, using a single USB camera. The core problem it addresses is what the README calls the "long lie", the dangerous gap between when someone falls and when help arrives. Existing approaches either send video to a cloud server (raising privacy concerns and requiring internet) or use models too computationally heavy for small devices. This project does neither: everything runs locally on the device, no frames leave it, and no cloud account is needed. The detection logic works by analyzing body posture in each camera frame. It uses a pose estimation model (YOLO11n-pose) that identifies 17 body landmarks per person in each frame, then tracks whether the body axis has shifted from vertical (standing) to horizontal (fallen) and whether the person has stayed still afterward. This is handled by a state machine with three states: normal, possible fall, and fall confirmed. A fall is only confirmed if the body goes horizontal rapidly and stays horizontal and still for a minimum duration, this filters out actions like sitting down or bending over. On the Jetson Nano with TensorRT FP16 optimization, the system runs at 22, 25 frames per second with an end-to-end latency of about 2 seconds from motion to alert. In testing on 60 video clips, it achieved 90% recall on fall clips and a 2.5% false-positive rate on non-fall actions. All detection thresholds are configurable without changing code. The tech stack is Python 3.10, Ultralytics for YOLO inference, OpenCV for camera handling, NumPy for geometry math, PyTorch as a dependency, and TensorRT for deployment optimization.