Robust K-means algorithm with automatically splitting and merging clusters and its applications for surveillance data

With the pervasive of the definition of the smart city, the data volume of the surveillance system, huge number of video surveillance devices is now rapidly expanding. The research to surveillance data mining and analytics has attracted increasing attention due to its applications. Cluster analysis as an important task of data mining in video surveillance has recently been highly explored. K-means algorithm is the most popular and widely-used partitional clustering algorithm in practice. However, traditional k-means algorithm suffers from sensitive initial selection of cluster centers, and it is not easy to specify the number of clusters in advance. In this paper, we propose a robust k-means algorithm that can automatically split and merge clusters which incorporates the new ideas in dealing with huge scale of video data. This novel algorithm not only addresses the sensitivity in selecting initial cluster centers, but also is resilient to the initial number of clusters. The performance is experimentally verified using synthetic and publicly available datasets. The experiments demonstrate the effectiveness and robustness of the proposed algorithm. Moreover, experiment is conducted on a real video surveillance dataset and the result shows that the novel approach can be applicated friendly in video surveillance.