Semantic Scene Modeling for Behavior Analysis Using LDA Mixture Model

摘要

Incomplete trajectory will result in over segmentation of semantic region and inappropriate region merging. In order to segmentation trajectory, CAIs rule, credibility assurance rule, non cross-regional rule and feature point's segmentation rule are introduced. A motion words generation method based on trajectory segments is then proposed. In the process of semantic scene modeling, firstly considers constraints of the behavior categories to establish topic space for all kinds of behavior; and then model the behavior context through hidden Markov chain participating in; finally a LDA mixture model is proposed. Experiments show that semantic scene modeling can access a good global semantic structure of the scene, and the behavior classification effect of LDA mixture model is more ideal.