A probabilistic model for camera zoom detection

Camera motion detection is essential for automated video analysis. We propose a new probabilistic model for detecting zoom-in/zoom-out operations. The model uses EM to estimate the probability of a zoom versus a non-zoom operation from standard MPEG motion vectors. Traditional methods usually set an empirical threshold after deriving parameters proportional to zoom, pan, rotate and tilt. In contrast, our probabilistic model has a solid probabilistic foundation and a clear, simple probability threshold. Experiments show that this probabilistic model significantly out-performs a baseline parametric method for zoom detection in both precision and recall.