Transformer-Enhanced Activity Recognition From Wearable Devices: A Current Analysis With Modern Datasets

Elderly people's ability to live independently is greatly aided by intelligent assistive technology that can identify Activities of Daily Living (ADLs). The potential of attention-based Transformer models for complicated ADL identification is still unexplored, despite the widespread use of deep learning models like CNNs and LSTMs for Human Activity identification (HAR) [2], [5]. Using the most recent multi-modal wearable dataset (Wearable-ADL-2023) [7], which contains information from smartphones, smartwatches, and inertial measurement units (IMUs), this study offers an improved implementation and evaluation of a hierarchical Transformer-based model for ADL classification. In order to classify seven composite ADLs, our model first uses a transfer learning strategy to identify atomic activities [4], then uses their probability scores as input to a Transformer. According to experimental results, the Transformer outperforms LSTM baselines (71.85%) in subject-dependent scenarios, achieving 78.42% accuracy [10]. It also has greater ability to capture long-range dependencies without repetition. We also address overfitting issues, segment length's effect on performance, and model resilience.