双峰分布の確率的モデリングに基づく運動物体の検出

概要

論文の詳細を見る
The motion analysis based on local methods allows one to avoid sophisticated modeling of object motion. Statistical models as well as labeling method are one of possible approaches to tackle the problem of motion, segmentation. Recently, the analysis of color RGB sequences has become popular. In our previous work [1] we used four six-variate Gaussian functions to describe the total motion RGB sequence. More precisely, the model from paper [1] described the motion in two consecutive frames; one Gaussin function is responsible for modeling background and three others are used for moving object. Such a model works well for many test sequences however, in some cases the quality of results is visibly lower or even unacceptable. To overcome those problems we decided to modified the model. In the past solution the use of three Gaussian functions for moving object resulted from assumption that the brightness of the object can be modeled by one Gaussian distribution. For some moving object whose histogram has two dominating maxima such a modeling simplification has been to "large". In our new model we decided to assume that the brightness of the moving object can be described by two Gaussian functions, i.e., some object areas, for example bright ones, are modeled by a Gaussian and others, for example dark ones, are modeled by another Gaussian. In other word, we assume a priori that there are two modes in the object histogram, i.e., bimodal moving object modeling.
社団法人電子情報通信学会の論文
1996-09-18

双峰分布の確率的モデリングに基づく運動物体の検出

スポンサーリンク

概要

著者

関連論文

スポンサーリンク