Motion2Vector: Unsupervised Learning in Human Activity Recognition Using Wrist-Sensing Data

With the increasing popularity of consumer wearable devices augmented with sensing capabilities (smart bands, smart watches), there is a significant focus on extracting meaningful information about human behaviour through large scale real-world wearable sensor data. The focus of this work is to develop techniques to detect human activities, utilising a large dataset of wearable data where no ground truth has been produced on the actual activities performed. We propose a deep learning variational auto encoder activity recognition model - Motion2Vector. The model is trained using large amounts of unlabelled human activity data to learn a representation of a time period of activity data. The learned activity representations can be mapped into an embedded activity space and grouped with regards to the nature of the activity type. In order to evaluate the proposed model, we have applied our method on public dataset - The Heterogeneity Human Activity Recognition (HHAR) dataset. The results showed that our method can achieve improved result over the HHAR dataset. In addition, we have collected our own lab-based activity dataset. Our experimental results show that our system achieves good accuracy in detecting such activities, and has the potential to provide additional insights in understanding the real-world activity in the situations where there is no ground truth available.