Real-Time Fall Recognition Using a Lightweight Convolution Neural Network Based on Millimeter-Wave Radar

Fall recognition is very important for the elderly. Consequently, fall recognition using convolution neural networks has been widely studied. However, current fall recognition studies rarely consider convolution neural network implementation on radar devices. As radar devices have limited processing power and storage space, recognition networks with high computational complexity, time-consuming, and parametric quantities will limit their terminal applications. Hence, fall recognition requires a lightweight convolutional neural network (LWCNN). With regard to this, an LWCNN comprised of channel shuffling (CS), grouped convolution, and residual approach is proposed for real-time fall recognition. We built a database of radar micro-Doppler signatures in two indoor environments, which includes data on falls, standing after falls, and similar falls and standing after falls. We trained and tested the proposed LWCNN in the constructed database to validate its performance. The test results show that our LWCNN minimizes parameters and floating-point operations (FLOPs) while achieving excellent recognition accuracy and shorter execution time compared to other LWCNNs. Besides, compared to state-of-the-art (SOTA) fall recognition systems using radar, our fall recognition system has good generalization ability to multiple classes of highly similar behaviors.