Unsupervised Techniques for Video Clustering, Classification and Understanding

SET Corporation proposes to design, develop and demonstrate unsupervised algorithms for automatic grouping, classification and characterization of prerecorded or live video streams. Our system integrates algorithms for color, appearance, and motion-based characterization of video segments of possibly variable lengths using unsupervised techniques that allow efficient user interactions. We propose a simultaneous segmentation and parameter learning approach that clusters video segments in an unsupervised manner. A sequence of such video segments forms a video cluster from which, radiometric, statistical, structural and semantic attributes are extracted. The segmentation algorithm is quasi-invariant to viewpoint changes and invariant to variations such as the execution rate of events, resolution of video, rotation and translation. In Phase II, we will integrate a high-level semantic architecture that will provide summarization, indexing and retrieval functionalities.