Automatic Dominant Motion Characterization in the Video

Dominant motion detection is essential for automated video analysis. Traditional methods usually set an empirical threshold to detect pan/tilt/zoom. In this paper, we propose a novel approach of automatic detecting video dominate motion which is parameter-free. Based on the motion trajectories of feature points, the distribution of motion is estimated by kernel density estimation. A Kullback-Leibler divergence based K-Nearest Neighbor classifier is used to classify the motion into pan/tilt/zoom etc category. Experiments results on both standard test video and real world video show this method significantly out-performs a baseline parametric method for dominate motion detection in both precision and recall.