Autonomous Learning of a Robust Background Model for Change Detection

We propose a framework for observing static scenes that can be used to detect unknown objects (i.e., left luggage or lost cargo) as well as objects that were removed or changed (i.e., theft or vandalism). The core of the method is a robust background model based on on-line AdaBoost which is able to adapt to a large variety of appearance changes (e.g., blinking lights, illumination changes). However, a natural scene contains foreground objects (e.g., persons, cars). Thus, a detector for these foreground objects is automatically trained and a tracker is initialized for two purposes: (1) to prevent that a foreground object is included into the background model and (2) to analyze the scene. For efficiency reasons it is important that all components of the framework are using the same efficient data structure. We demonstrate and evaluate the developed method on the PETS 2006 sequences as well as on own sequences of surveillance cameras.