SFU at TRECVid 2010: Surveillance Event Detection.
Zhi feng Huang, Greg Mori · 2010
This paper describes the SFU entry for the TRECVid 2010 surveillance event detection challenge. We focus on detecting the ”PresonRuns ” event. The system has three stages. First, a pre-processing step is performed to extract moving candidate space-time regions using background subtraction, optical flow, and photogrammetric context. A short clip representation is constructed for each moving candidate region with a human figure detection. Motion features are extracted from the clip representation, and the features are scored by an AdaBoost classifier. Finally, a post-processing step produces temporally localized responses for evaluation. 1.