Combined feature-level video indexing using block-based motion estimation

We describe a method for attaching content-based labels to video data using a weighted combination of low-level features (such as colour, texture, motion, etc.) estimated during motion analysis. Every frame of a video sequence is modeled using a fixed set of low-level feature attributes together with a set of corresponding weights using a block-based motion estimation technique. Indexing a new video involves an alternative scheme in which the weights of the features are first estimated and then classification is performed to determine the label corresponding to the video. A hierarchical architecture of increasingly complexity is used to achieve robust indexing of new videos. We explore the effect of different model parameters on performance and prove that the proposed method is effective using publicly available datasets.