Visual-HAR: A Real-Time Human Activity Recognition Model Based on Visual Sensors

The paper presents a human activity recognition model, Visual-HAR, which leverages visual sensors to identify activities of elderly individuals. The goal of Visual-HAR is to accurately and promptly detect whether an elderly person is in danger by recognizing their activities in real time. This model utilizes convolutional neural networks (CNNs) and focuses on improving the ConViVit model through hyperparameter optimization. The performance of both Visual-HAR and Con-ViVit models was evaluated using three human action datasets: NTU RGB+D60, PKU-MMD, and PKU-MMD2. The evaluation criteria included both accuracy and execution time. The results indicated that Visual-HAR significantly outperformed ConViVit with an average testing performance improvement of 7.41% and a reduction in execution time by at least 36%, which is critical for real-time decision-making. These findings highlight Visual-HAR's potential as a reliable tool for monitoring elderly individuals.