Human Activity Recognition on Smartphones using Symbolic Data Representation

In ubiquitous computing, Human Activity Recognition (HAR) systems have an important role to enabled continuous monitoring of human behavior. This technology can be useful in healthcare applications, for monitoring patients' health and encourage a healthy lifestyle. In this paper, we focus on features extraction stage of a HAR system. Many studies for mobile and wearable sensor-based HAR have applied manually engineered features that need domain expert knowledge. However, trust on such knowledge is problematic when aiming to generalize across different application domains. To overcome this problem, we present a novel approach for HAR based on symbolic data representation of time series that extract structural features without human efforts. The Bag-Of-SFA-Symbols (BOSS) method is extended to multi-dimensional time series, in order to enable that symbolic representation can be used to process the inertial sensors data. A comparative study between the proposed method and four machine learning classifiers with handcraft features is presented. Experiments on accelerometer data from three publicly datasets were executed for subject-dependent and subject-independent evaluation. The results show that our method achieves good accuracy performace across datasets and aplications, and substantial recognition improvement over a baseline.