Abrupt Shot Change Detection Using Multiple Features and Classification Tree

Abstract

We propose an abrupt shot change detection method using multiple features and classification tree. Typical shot change detection algorithms have usually used single feature obtained between consecutive frames, and the shot change is determined with only one fixed threshold in whole video sequences. However, the contents of the video frames at shot changes such as intensity, color, shape, background, and texture change simultaneously. Thus multiple features have the advantage of single feature to detect shot changes. In this paper, we use five different features such as pixel difference, global and local histogram difference, and block-based difference. To classify the shot changes with multiple features, we use the binary classification tree method. According to the result of classification, we extract important features of the multiple features and obtain threshold value for feature at each node of the tree. We also perform the cross-validation analysis and drop-case method to confirm the reliability of the classification tree. An experimental result shows that our method has better performance than the existing single feature method for detecting abrupt shot changes.