EVENT recognition总结

	activity recognition	abnormal event recognition
model-based	HMM, coupled HMM ,and Dynamic Bayesian Network	semisupervised adapted(HMM) framework	these techniques rely on the choice of good models, which requires sufficient training data to learn the model parameters.
Appearance-based techniques	1volumetric features based on optical flow representations 2volumetric features from regions with local variations in both spatial and temporal dimensions 3detect salient regions by extending the idea of Harris interest point operators and  applied separable linear filters for the same objective	extracting an ensemble of densely sampled local video patches	reliable extraction of spatial-temporal features and/or salient regions, which are often based on optical flow or intensity gradients in the temporal dimension. sensitive to motion，detected interest regions are often associated with high motion regions