Human activity recognition based on IMU sensors using a combination of convolutional neural network and multi-head attention

Human activity recognition (HAR) and fall detection play crucial roles in healthcare for the elderly and remote observation. This research presents three innovative deep learning models - MSRLSTM, MSRLSTM - Refined, and MSR - MultiHeadAttention-designed for HAR and fall detection by utilizing Inertial Measurement Unit data from the UP-Fall Detection Dataset. By employing convolutional neural networks, residual learning, and multi-head attention, these models effectively capture complex temporal and spatial patterns present in multimodal sensor data. When evaluated on the UP-Fall Detection Dataset, MSRLSTM-Refined and MSR - MultiHeadAttention achieved accuracies of 93.91% and 95.49%, respectively, outpacing the baseline MSRLSTM (92.10%). The MSR-MultiHeadAttention model stands out due to its precision and temporal modeling capabilities, while MSRLSTM-Refined delivers computational efficiency suitable for wearable devices. Although there are challenges in differentiating similar motion patterns, these models demonstrate significant potential for real-time fall detection, contributing to remote healthcare monitoring solutions and related fields.