Method for scene modeling and change detection
View Patent ↗A method for dynamic scene modeling and change detection applicable to motion analysis utilizes optical flow for capturing and modeling the dynamics of the scene. Uncertainties in the measurements are evaluated and utilized in order to develop a robust representation of the scene in a higher dimensional space. In another embodiment, a dynamical model of the scene is developed that utilizes multiple past frames to predict the next frame. Incremental methods for updating the model are developed and, towards detection of events, a new measure is introduced that is based on a state-driven comparison between the prediction and the actual observation.
1. A method for scene modeling and change detection in an image, comprising the steps of:
computing optical flow for said image;
performing an invariant transformation such that image pixel intensity is transformed and evaluated in an illumination-invariant space;
forming a background model in a highdimensional space;
utilizing results of said computing optical flow and of said invariant transformation as features in said background model;
utilizing said background model to estimate probability for a current input to belong to the background;
providing a first and a second indication whenever said probability is respectively above and below a given threshold;
adding said current input to said image background model whenever said probability is above said threshold;
adding said current input to said image background model with a low probability whenever said probability is below said threshold; and
performing morphological operations on said pixel-level detection for outputting detection.
2. A method as itcited in claim 1 , wherein said background model is developed in a high-dimensional space using kernel density estimation.
3. Apparatus for scene modeling and change detection in an image, comprising:
means for computing optical flow for said image;
means for performing an invariant transformation such that image pixel intensity is transformed and evaluated in an illumination-invariant space;
means for forming a background model in a high-dimensional space;
means for utilizing results of said computing optical flow and of said invariant transformation as features in said background model;
means for utilizing said background model to estimate probability for a current input to belong to the background;
means for providing a first and a second indication whenever said probability is respectively above and below a given threshold;
means for adding said current input to said image background model whenever said probability is above said threshold;
means for adding said current input to said image background model with a low probability whenever said probability is below said threshold; and
means for performing morphological operations on said pixel-level detection for outputting detection.