Data Generation Process Modeling for Activity Recognition

Abstract

The dynamics of body movements are often driven by large and intricate low-level interactions involving various body parts. These dynamics are part of an underlying data generation process. Incorporating the data generation process into data-driven activity recognition systems has the potential to enhance their robustness and data-efficiency. In this paper, we propose to model the underlying data generation process and use it to constrain training of simpler learning models via sample selection. As deriving such models using human expertise is hard, we propose to frame this task as a large-scale exploration of architectures in charge of relating sensory information coming from the data sources. We report on experiments conducted on the Sussex-Huawei locomotion dataset featuring a sensor-rich environment in real-life settings. The derived model is found to be consistent with existing domain knowledge. Compared to the basic setting, our approach achieves up to 17.84% improvement, by simultaneously reducing the number of required data sources by one-half. Promising results open perspectives for deploying more robust and data-efficient learning models.