A Microtask-Based Segmentation Method for Acceleration Data for Human Activity Recognition

This paper proposes a microtask-based segmentation method for acceleration data for human activity recognition. The proposed method finely segments unlabeled acceleration data with a sliding window, clusters using cepstral coefficients and variance for each window, and assigns cluster numbers to the windows. When the cluster numbers of two adjacent windows dif-fer, a tentative segmentation is performed by placing a boundary between the windows. Subsequently, the validity of the tentative segments is examined through surveys from unspecified large numbers of people, consolidating or splitting segments as needed. In evaluation experiments with an allowable error of 30 seconds, for the PAMAP2 dataset, tentative segmentation achieved 1,000 recall and 0.365 precision, while both recall and precision reached 1.000 after human questionnaire surveys. For the WEAR dataset, tentative segmentations achieved 1.000 recall and 0.775 precision, while recall and precision reached 0.707 and 0.909, respectively after human questionnaire surveys.