Human activity classification in people centric sensing exploiting sparseness measurement

Existing human activity recognition models in people centric sensing have explored different features of the mobile phone data to achieve considerable accuracy. However, the heterogeneity of such data caused by different phone carrying modes during data collection process remains a challenge. It has been observed by us that although the waveform of tri-axial accelerometer data varies in different modes, its sparseness within segmented frames tends to preserve in general. In this paper, we propose to adopt an augmented feature set by taking into account the sparseness measurement to improve the robustness of human activity classification. It has been shown in the experiment results that the sparseness features have a high importance ranking among the list. In addition, the experiment results show that the AUC measurement results of the Random Forest model can be improved both in single and mixed phone carrying modes, which justify the effectiveness of the sparseness measure in addressing the heterogeneity of tri-axial accelerometer data on mobile phones.