Video event detection based on over-segmented STV regions

Real-world environment introduces many variations into video recordings such as changing illumination and object dynamics. In this paper, a technique for abstracting useful spatio-temporal features from graph-based segmentation operations has been proposed. A spatio-temporal volume (STV)-based shape matching algorithm is then devised by using the intersection theory to facilitate the definition and detection of video events. To maintain system efficiency, this research has integrated an innovative feature-weight evaluation mechanism which “rewards” or “punishes” recognition outputs based on the segmentation quality. Substantial improvements on both the event “Precision” and “Recall” rate and the processing efficiency have been observed in the experiments in the project.