AT&T Research at TRECVID 2013 : Surveillance Event Detection

T his paper present s a gener al surveillance event detection system that participates in the interactive Surveillance Event Detection ( i SED) task of TRECVID 2013 . In the proposed system, a set of spatio - temporal features including Space - Time Interest Points (STIP) and Dense Trajectories are extracted, and a sliding temporal window is empl oyed as the detection unit. Fisher Vector is used to encode lo w - level features as the representation of each sliding window. Both feature - level and decision - level fusion s are used to combine multiple features . In order to deal with the highly imbalanced nature of surveillance data, the system performs detections usin g the CascadeSVMs algorithm according to each specific event and camera view . Two different interactive environments are evaluated, one focuses on high throughput and the other includes related result expansion. In the primary run evaluations, o ur system r anks the top in 2 out of 7 event detection tasks.