On-Line Unsupervised Segmentation for Multidimensional Time-Series Data and Application to Spatiotemporal Gesture data

摘要

This paper proposes an on-line unsupervised segmentation algorithm of multidimensional time-series data. On-line unsupervised segmentation algorithm is important for discovery of frequent patterns from a large amount of multidimensional sensor data such as whole motion data and multi-modal data, because real-world problems are often dynamic and incremental:the input data may change over time and on-line segmentation methods are significantly faster than batch methods. We enhance the Sliding Window and Bottom-up (SWAB) algorithm toward the implementation of the segmentation algorithm of multidimensional data and propose MD-SWAB (Multi Dimensional SWAB). To evaluate the proposed algorithm, we use the continuous hand gesture data observed with a motion capture system. We have found that MD-SWAB outperforms a comparative algorithm in the segmentation performance and is significantly faster than the existing algorithm. In addition, we combine it with the on-line clustering method HB-SOINN to label the detected segments. The result of experiment shows that gesture patterns which are belong to same category are represented as similar label sequences.