Comparing LSTM and GRU for Fall Detection in FMCW Radar-Based Point Cloud Data

In safety monitoring systems, automatic fall detection is a major problem, particularly for high-risk and elderly people. There are often issues with camera-based techniques, such as privacy issues and lighting disturbances, which call for more reliable alternatives. This study proposes using Frequency-Modulated Continuous Wave (FMCW) radar as a non-visual sensor to detect human activity through point cloud data representing XYZ position, velocity, and acceleration in time-series form. The method used is Recurrent Neural Network (RNN), specifically Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), with a focus on tuning parameters such as the number of timesteps, memory units, activation functions, and optimizers. Experimental results show that GRU with 30 timesteps, tanh activation, and Adam optimizer produces the best performance with a training accuracy of 98.89% and a validation accuracy of 98.92%, outperforming LSTM in similar configurations. In addition, GRU shows higher recall in detecting falls, making it an efficient and accurate model choice. These findings indicate that radar point clouds can be effectively used in a precise real-time fall detection system that does not compromise user privacy.