Efficient Video Event Retrieval using Ensemble Subspace Techniques (EVEREST)

As the use of persistent aerial video surveillance has grown, the means to produce imagery has far outstripped the means to exploit imagery. Due to this discrepancy, video is often archived for off-line review, resulting in massive video databases that are drastically under-utilized today. To help analysts locate valuable intelligence residing in these databases, we propose a system for Efficient Video Event Retrieval using Ensemble Subspace Techniques, or EVEREST. The system automatically extracts from video a hierarchical set of descriptors encoding several distinct levels of semantic concepts. EVEREST constructs an ensemble of randomized indexes over descriptor subspaces, called a random decision forest (RDF). In contrast to the difficulties encountered by most other indexing techniques in high-dimensionality spaces, our RDF-based indexing method effectively exploits the large number of dimensions by executing database searches and updates in parallel across the ensemble of decision trees. To ensure scalability, we will leverage modern commodity multi-processor architectures. EVEREST’s interface allows analysts to query the database for video that exhibits similarity to a given exemplar video, where similarity is interactively defined by selecting attributes of interest about the video. Search results may be iteratively refined by flagging positive and negative matches and re-executing the query.