Abstract

This paper describes the SFU entry for the TRECVid 2009 event detection challenge. We attempted to detect events involving either one or two individuals, and present algorithms for “running”, “embrace”, and “pointing” event detection. The system has three stages. First, preprocessing is performed to extract candidate space-time regions using background subtraction and photogrammetric context. These candidate regions are processed with specific event detection algorithms, utilizing motion for “running” and “embrace ” and shape for “pointing”. A postprocessing step produces temporally localized responses for evaluation. The system is computationally efficient (0.2 seconds per frame) and achieved very good minimum DCR scores on the final evaluation. 1.

Citations