Robust Motion Detection with Temporal Decomposition and Statistical Regularization
Abstract
This paper deals with the detection of moving objects. We have defined a method able to cope with perturbations frequently encountered during acquisition of outdoor image sequences: camera not perfectly stationary, illumination modifications, occlusions, ... Temporal integration and statistical regulariz- ation are the two main features of the method. A temporal multiscale decomposition allows us to detect and to characterize various dynamical behaviours of the elements present in the scene. A tracking module provides a prediction map, which gives a confidence level for presence of motion at a given pixel. A statistical regularization framework, based on Markov random field models, supplies a formal way to combine these different sets of computed information, while exploiting a priori knowledge on the primitives to be determined. A calibration technique based on so-called qualitative boxes is used to estimate model parameters. Several experiments with real image sequences depicting various complex situations have validated the approach.