Auroral Video Temporal Segmentation Based on Time-constrained Spectral Clustering
Qian Wang, XiangWei Hou, JiuLing Du · 2016
Acquiring a large amount of samples is a prerequisite for statistical analysis. In study of auroral phenomena, a tool which is able to segment 24-hour uninterrupted auroral observation into subsequences is urgently provided. This study aims at segmenting long auroral video into isolate auroral events based on clustering algorithm. The primary problem is how many events exist in a long uninterrupted auroral observation, i.e. deciding the number of clusters. We proposed a cluster validity index dedicated for manifold structure considering consistency within clusters and regional distribution between clusters. Then, we used the spectral clustering on the auroral image sequence. In order to guarantee the continuity and isolation of segments, a new temporal constrain was incorporated into clustering algorithm. Using the proposed method, the auroral videos were automatically segmented into a proper number of events. Compared with keogram, the auroral appearance and motion in the resultant segments exhibit homogeneity and the number of segments conform to the actual situation.