Detection of activities and events without explicit categorization

摘要: We address the problem of unsupervised detection events (e.g., changes or meaningful states human activities) without any similarity test against specific models probability density estimation category learning). Rather than estimating densities, very difficult to calculate in general settings, we formulate event as binary classification with ratio [9] a hierarchical probabilistic framework. The proposed method takes pairs video stream data (i.e., past and current) input differing time-scales, generates way online learning, judges if there is ‘meaningful difference’ between them based on multiple estimations. Through experimental studies real-world scenes domains using challenging datasets from sports scene tennis match) complex background, demonstrate potential advantage our approach over state-of-the-art terms precision efficiency.