Abstract:

The activity pattern mining technique is  the  key component of semantic analysis for surveillance video. Because of the lack of prior and highdimensional feature constraints, the complexity of the model structure of the parametric mining model is  difficult to be precisely defined. Nonparametric clustering of motion feature by infinite Gaussian mixture was used to get the elementary activity patterns, based on which duration distribution was estimated. The  partial-dimension test for feature validated the motion similarity hypothesis existing  in the mining model. The results showed that the obtained activity patterns precisely reflected motion semantics of the scene, and that the multi-modality temporal distribution existing  in activity can  be further used to discover the hidden knowledge of motion.

Key words: visual surveillance, scene analysis, infinite Gaussian mixture model, activity pattern mining