Fall Detection via Inaudible Acoustic Sensing
summary
System emits a 20 kHz continuous wave from a home audio speaker and records Doppler returns on a microphone, denoises via spectral subtraction and STFT, then locates motion with a power-burst curve and extracts torso/leg/arm speed, extreme-frequency-ratio, and spectral-entropy features. Features reduce with SVD and K-means and feed a Hidden Markov Model. Evaluation with 23 volunteers across four rooms yields >90% metrics for falls vs fall-like and object-drop events.
AI-written summary · about