Abstract: Domain motion analysis is an important research area in content based video index and retrieve. We propose a novel approach of automatic detecting video domain motion which doesn't rely on threshold. Based on the motion vectors of feature points, the distribution of motion is estimated by kernel density estimation. A Kullback-Leibler divergence based k-Nearest Neighbor classifier is proposed to classify the camera moving into pan八ilt/zoom etc category.The training samples arc generated automatic according to the character of each domain motion. Experimental results show that our proposed method is capable of achieving both accuracy and robustness.

Key words: Motion classification,Kernel density estimate,kNN