AttriNet: learning mid-level features for human activity recognition with deep belief networks

Human activity recognition (HAR) is essential to many context-aware applications in mobile and ubiquitous computing. A human's physical activity can be decomposed into a sequence of simple actions or body movements, corresponding to what we denote as mid-level features. Such mid-level features ("leg up," 'leg down," "leg still,"...), which we contrast to high-level activities ("walking," "sitting,"...) and low-level features (raw sensor readings), can be developed manually. While proven to be effective, this manual approach is not scalable and relies heavily on human domain expertise. In this paper, we address this limitation by proposing a machine learning method, AttriNet, based on deep belief networks. Our AttriNet method automatically constructs mid-level features and outperforms baseline approaches. Interestingly, we show in experiments that some of the features learned by AttriNet highly correlate with manually defined features. This result demonstrates the potential of using deep learning techniques for learning mid-level features that are semantically meaningful, as a replacement to handcrafted features. Generally, this empirical finding provides an improved understanding of deep learning methods for HAR.