Application of Transformer to Wearable Sensor-based Complex Human Activity Recognition

Deep learning technologies such as convolutional neural networks and long short-term memory have been approved to be effective in sensor-based human activity recognition. However, the existing works using these models have limitations on applications due to their large volume of computations. This paper presents an effective sensor-based human activity recognition model based on Transformer which provides an improved accuracy for sensor-based human activities recognition. In this work, we modified the existing Transformer models for human activity recognition to simplify the attention mechanism and reduce computation complexity. Furthermore, unlike the current works on sensor-based activity recognition which only tested simplex activities, this work targeted a complex and larger set of human activities and the results demonstrated the effectiveness of the proposed work.