FPGA Implementation of a Radar-Based Fall Detection System Using Binarized Convolutional Neural Networks

Area of Science:

• Electrical Engineering
• Computer Science
• Gerontology

Background:

• Increasing elderly population living alone necessitates effective fall detection systems.
• Existing systems face challenges with privacy, line-of-sight, continuous monitoring, and high computational complexity.
• Need for compact, low-power, and efficient hardware for distributed fall detection.

Purpose of the Study:

• To propose a lightweight fall detection system using continuous-wave (CW) radar and a binarized convolutional neural network (BCNN).
• To achieve efficient hardware implementation for reduced power consumption and smaller hardware footprint.
• To validate the system's accuracy and processing speed for real-time fall detection.

Main Methods:

• Utilized continuous-wave (CW) radar sensors for non-invasive monitoring.
• Applied short-time Fourier transform (STFT) for preprocessing radar signals into binary spectrograms.
• Developed a binarized convolutional neural network (BCNN) for activity classification.
• Implemented preprocessing and classification modules as hardware accelerators on a field-programmable gate array (FPGA) within a system-on-chip (SoC) architecture.

Main Results:

• Achieved 96.1% accuracy in classifying five fall activities and seven non-fall activities.
• Hardware acceleration resulted in significant speedups: 387.5× for preprocessing and 86.7× for classification compared to software.
• Overall system processing time reduced to 2.58 ms, an 89.5× speedup over the software baseline.

Conclusions:

• The proposed lightweight CW radar and BCNN system offers an efficient and accurate solution for fall detection.
• Hardware acceleration on FPGA significantly enhances processing speed and reduces power consumption, making it suitable for distributed deployment.
• This technology holds promise for improving safety and independence for the elderly living alone.