A unified approach to background adaptation and initialization in public scenes

Foreground detection methods generally assume that backgrounds are observed more frequently than foregrounds are, but the assumption is not valid in public scenes. Viewing background adaptation in public scenes as a unified problem with background initialization and stationary object detection, we formulate it as an energy minimization problem in dynamic Markov random fields. Constraining the connections among the sites with spatiotemporal reliabilities, we robustly handle object-wise changes and efficiently minimize the energy terms with a coordinate descent method. Evaluated with realistic sequences from i-LIDS, PETS, ETISEO and changedetection.net datasets, the proposed method outperforms state-of-the-art methods and temporal parameter adjustment.