Pattern discovery in motion time series via structure-based spectral clustering

Xiaozhe Wang, Liang Wang, Anthony Wirth

Abstract: This paper proposes an approach called ‘structure-based spectral clustering’ to identify clusters in motion time series for sequential pattern discovery. The proposed approach deploys a ‘statistical feature-based distance computation’ for spectral clustering algorithm. Compared to traditional spectral clustering approaches, in which the similarity matrix is constructed from the original data points by applying some similarity functions, the proposed approach builds the matrix based on a finite set of feature vectors. When the proposed approach uses less data points and simpler similarity function to computing the similarity matrix input for spectral clustering, it can improve the computational efficiency in constructing the similarity graph in spectral clustering compared to conventional approach. Promising experimental results with high accuracy on real world data sets demonstrate the capability and effectiveness of the proposed approach for pattern discovery in motion video sequences.