Tracking of Objects in Video Scenes with Time Varying Content

We propose a method for tracking of objects contained in video sequences. Each video object is represented by a set of polygonal regions. A bottom up approach (spatial segmentation/motion estimation) is applied for the initialisation of the method, a limited human interaction is used to build the semantic map of the first frame in video sequence. The tracking of this model along a video sequence is based on detecting and indexing new objects in a video scene. Semantic rules are used to label new objects and, the current state of segmentation is validated by forward projection of the background.