A unified shot boundary detection method based on linear prediction with Bayesian cost functions

The detection of edits in a video sequence is the first step for video analysis, which segments a video into its basic components. In this paper, we propose a novel and efficient approach for shot boundary detection, which can detect cuts and dissolves reliably using a uniform framework. Our approach is based on the temporal linear prediction of the frames. With frames in a video shot, a frame can be predicted from its previous frames. If the prediction error is high, a cut should happen. For dissolves, the gradual transitions can be modeled by temporal linear prediction with constant prediction coefficients. Experimental results show that our algorithm can achieve high precision even if a video contains object motion and camera motion, and is able to detect and classify cuts and dissolves in real time.