Automatic video similarity measure using vectorized attributed graph matching

Abstract

The video scene similarity is introduced in term of the objects' positions analysis. The background objects' positions which are extracted from the video scene are transformed into the spatial logical functions. The spatial logical function which is used to represent the objects' positions on the scene are decoded into top-view model which is projected on the grid unit plane later. From the simulated grid unit plane, the video scene top-view model information are kept in digital form which can be used variously and easily. Besides those data in grid unit plane are the features in the semantic spatial graph. The transformation from unit plane into graph is described in this dissertation. The semantic spatial graph is used to identify the pattern of the objects' position in a scene. Due to the characteristic of semantic spatial graph, the comparing between video scenes are simplified. The transformation to top-view model is robust to perspective. Eventhough the view are changed perspectively, the semantic spatial graph can determined the similarity through top-view model. The robustness of this algorithm are showed by the some examples of the experiments.