4 August 2023 in IEEE Sensors Journal
Fall detection has become a popular research field in recent years due to the fact that falls can cause serious injuries. Most existing fall detection systems rely on wearable devices or image recognition technology. However, these approaches possess limitations. For instance, most people take off wearable devices during showers, while cameras with image recognition raise privacy concerns. To overcome these challenges, this study proposes a novel approach using 60 GHz radar to develop a fall detection system. We collected raw data from the frequency-modulated continuous wave (FMCW) radar in three bathrooms and two other indoor fields. Then, we derived features by generating Range-Doppler images and proposing a new Doppler image preprocessing method, Long-Time Interval Range-Doppler Histogram, each histogram represents the cumulative speed information of the Doppler images. Specifically, each histogram is composed of 30 Doppler images per second, capturing the overall speed distribution over a longtime interval. In addition, we proposed two deep learning models, a trigger model, and a fall detection model, for the fall detection system. The trigger model employs convolutional neural networks (CNNs) to identify Range-Doppler images with significant speed and distance variations and is optimized to run on devices with limited central processing unit (CPU) resources. The fall detection model utilizes a bidirectional long short-term memory-based deep learning model to make the final decision. Over 700 records were collected during the experiment, and the fall detection system achieved an accuracy rate of nearly 96%, with recall and precision rates exceeding 90%.