出版社：Information and Media Technologies Editorial Board
摘要：We propose a method of unsupervised event detection from a video that compares probability distributions of past and current video sequence data in a sequential and hierarchical way. Because estimation of probability distributions is known to be difficult, naively comparing probability distributions via probability distribution estimation tends to be unreliable in practice. To cope with this problem, we use the state-of-the-art machine learning technique called density ratio estimation : The ratio of probability densities is directly estimated without density estimation, and thus probability distributions can be compared in a reliable way. Through experiments on a walking scene and a tennis match, we demonstrate the usefulness of the proposed approach.
关键词：event detection;direct density-ratio estimation;cubic higher-order local auto-correlation
