Real-Time Fall Detection with Hybrid CNN-LSTM Using IWR6843AOP FMCW Radar

This paper presents a real-time fall detection system that utilizes a hybrid deep learning model combining one-dimensional Convolutional Neural Networks (1D CNN) and Long Short-Term Memory (LSTM) networks, integrated with FMCW IWR6843AOP radar sensor. The radar captures 3D point cloud data and extracts key kinematic features—position, velocity, and acceleration— across all spatial axes to provide compact yet informative input to the model. To ensure real-time performance, a multithreaded processing pipeline is implemented, allowing parallel execution of data acquisition, preprocessing, and inference tasks with minimal latency. Experimental results demonstrate that the selected model configuration with MaxPooling and Dropout achieves a training accuracy of 99.5%, while live tests across six previously unseen subjects yield a real-time fall detection accuracy of 96.7% with an average response delay of 0.60 seconds. These results confirm that the proposed system provides fast, accurate, and reliable fall detection suitable for real-world applications such as elderly care and ambient assisted living.