Human Activity Recognition Based on Transformer via Smart-phone Sensors
Liang Yu, Kaile Feng, Zizhuo Ren · 2023
Capturing the spatial and temporal relationships of time-series signals is a significant obstacle for human activity recognition based on wearable devices. Traditional artificial intelligence algorithms cannot handle it well, with convolution-based models focusing on local feature extraction and recurrent networks lacking consideration of the spatial domain. This paper offers a deep learning architecture based on transformer to address the aforementioned issue with data collected from smart-phones embedded with three-axis accelerometers. The transformer model, as a deep learning network mainly applied to natural language processing (NLP), is good at processing time-series information, where the self-attention mechanism captures the dependencies of perceptual signals in the temporal and spatial domains, improving the overall comprehensibility. We implement convolutional neural networks (CNN) and long and short-term memory networks (LSTM) for evaluation while our proposed model achieves an average classification accuracy of 94.84%, which is an improvement compared to the traditional model.