Efficient Recognition of Complex Human Activities Based on Smartwatch Sensors Using Deep Pyramidal Residual Network

Human activity recognition (HAR) has become a hot topic in artificial intelligence research due to the rapid development of smart wearable technologies. The goal of HAR is to accurately identify human actions using various data sources, such as video, images, and sensor data from wearable devices. Recent research in HAR has achieved promising results using learning-based methods, especially deep learning techniques. However, achieving state-of-the-art results remains a challenge for researchers. This study proposes a new approach to HAR that uses deep learning to classify human activities from smartwatch sensor data. We propose the use of a one-dimensional deep pyramidal residual network (1D-PyramidNet) for accurate human action identification. We evaluate the performance of our model against baseline models using the DHA dataset, a benchmark dataset for HAR that includes wristwatch sensor data for 11 complex human activities. The experimental results show that our 1D-PyramidNet model outperforms the baseline models, including CNN, LSTM, BiLSTM, GRU, and BiGRU. This confirms that the use of 1D-PyramidNet can improve the identification capabilities of HAR systems, achieving a maximum accuracy of 96.64%.