Transformer Model for Human Activity Recognition Using IoT Wearables

Human activity recognition plays a vital role in the modern life of its immense applications in medical care, sports, detection and prediction of unpredicted events and in biometrics field. HAR is the field of study for recognizing human activities from time-series data collected from video, images or from multimodal sensors equipped with smartphone, IoT wearables. A wide area of research in the field of HAR has been presented till now from machine learning and deep learning. Deep learning methods utilize the temporal view of the data, while shallow methods utilize the handcrafted features, statistical view. Most of the deep learning methods utilized CNN, LSTM architecture. Now its time to use the power of transformers in this area. For sequence analysis tasks, transformers were shown to outperform the deep learning methods. Transformers are proved to be best in finding relevant time steps and has the potential to model long term dependencies. In this work proposed, a novel method that employs a transformer-based model to classify human activities from data collected using Intertial measurement Units (IMU). This approach works by leveraging self-attention mechanisms in transformers to explore complex patterns from time-series data. In this paper, a transformer-based approach is experimented for HAR and its evaluation parameters are presented for two data sets USCHAD and WISDM. From the results, we can analyse the improvement in performance matrix and efficiency of the transformer-based model.