Real-Time Fall Detection Using Wideband Radar and a Lightweight Deep Learning Network

Radar-based fall detection in the state-of-the-art methods typically involves training on fixed long-duration data segments without considering action transitions, which does not align with practical work scenarios. In this article, a radar-based fall detection scheme that utilizes data streaming and lightweight networks is proposed to increase the accuracy of fall detection. An adaptive clutter suppression method of morphology is proposed to mitigate clutter including ghost targets from range-time spectrograms. A lightweight network is designed to detect falls in real time. Additionally, training samples with various radar heights and distances from subjects to wideband radar are expanded to establish our multi-indoor-scene behavior dataset. The proposed scheme obtained an accuracy of 0.9963 in a new scene with unseen subjects when the proposed network has a small size of 1.9178 MB. The experimental results demonstrate that increasing the diversity of training samples can improve fall prediction performance, and our method achieves better classification performance and stronger generalizability than the current state of the art in real-time fall detection using wideband radar.