Video retrieval by context-based interpretation of time-to-collision descriptors

Video retrieval using high-level indices is more meaningful than querying using low-level features. In this paper, we show how perceptual features such as time-to-collision (TTC) can lead to several high-level categories. Experiments hive been conducted to validate our proposed TTC detection algorithm to compute TTC from the divergence of the image velocity field. A simple and novel method named as the pilot cue is used to further refine our algorithm. Our initial system works with a rule-based approach where the extracted TTC shots (low-level feature) are mapped to their corresponding high-level indices. The information conveyed by their neighboring frames or shots (i.e. contextual information) is used to facilitate the mapping process. Several psychological effects (high-level indices) such as intimacy, suspense and terror are recovered as a result.