Smart-Wearable Sensors and CNN-BiGRU Model: A Powerful Combination for Human Activity Recognition

Human activity recognition (HAR) is a key component of ambient-assisted living and one of the most active areas of research in the Internet of Things (IoT). The use of wearable and embedded sensors in HAR overcomes the limitations of conventional approaches relying on machine vision and environmental sensors. We offer a novel, lightweight convolutional neural network–bidirectional gated recurrent unit (CNN-BiGRU) model that classifies human activities using the inertial sensor data collected with body-mounted smart-watches and smartphones. Unlike the traditional approaches, the presented model is trained on the magnitude of the 3-D acceleration ( $\widehat {\text {mag}_{a}}$ ), which significantly minimizes the input space 1-D. The deep learner has been validated using two different publicly available datasets from the wireless sensor data mining (WISDM) lab and different evaluation parameters, such as recall/sensitivity, precision, accuracy, and F1-scores, are computed. A comparison with the existing studies reveals that our proposed learner surpasses the existing methodologies. Using magnitude of 3-D acceleration ( $\widehat {\text {mag}_{a}^{w}}$ , 1-D input signal), we have achieved 97.29% accuracy for all six activities of WISDM 2011 dataset and 98.81% accuracy for 3-D acceleration ( ${a}_{x}^{w},{a}_{y}^{w}$ , and ${a}_{z}^{w}$ ). The precision, recall, and F1-score remained at 97% for the 1-D case and 99% for the 3-D case. When evaluated on the data of all 18 smartwatch-based activities in the WISDM 2019 dataset, we have achieved 97.5% accuracy with the magnitude of 3-D acceleration ( $\widehat {\text {mag}_{a}^{w}}$ ) and 98.4% accuracy for 6-D acceleration and angular velocities ( ${a}_{x}^{w},{a}_{y}^{w},{a}_{z}^{w},\omega _{x}^{w},\omega _{y}^{w}$ , and $\omega _{z}^{w}$ ). The precision, recall, and F1-score remained at 98% in both cases.