Deep feature learning and selection for activity recognition

Human physical activity recognition from sensor data is a growing area of research due to the widespread adoption of sensor-rich wearable and smart devices. The growing interest resulted in several formulations with multiple proposals for each of them. This paper is interested in activity recognition from short sequences of sensor readings. Traditionally, solutions to this problem have relied on handcrafted features and feature selection from large predefined feature sets. More recently, deep methods have been employed to provide an end-to-end classification system for activity recognition with higher accuracy at the expense of much slower performance. This paper proposes a middle ground in which a deep neural architecture is employed for feature learning followed by traditional feature selection and classification. This approach is shown to outperform state-of-the-art systems on six out of seven experiments using publicly available datasets.