MediaCCNY at TRECVID 2012: Surveillance Event Detection.
Xiaodong Yang, Chucai Yi, Liangliang Cao, Yingli Tian · TRECVID · 2012
In this paper, we present a general event detection system evaluated by the Surveillance Event Detection (SED) task of TRECVID 2012 campaign. The proposed system is evaluated on all the seven event categories of the SED task. In our system, a sliding temporal window is employed as the detection unit, which is represented by a histogram of spatio-temporal features including STIP-HOG/HOF and SURF/MHI-HOG. We also investigate the spatial priors of various events by estimating spatial distributions of actions under different camera views in the training data. As non-linear SVMs usually have superior performances but in general are much slower in both training and testing, we therefore employ explicit feature maps to approximate large scale non-linear SVMs by linear ones. In order to deal with highly imbalanced data, our system performs detections by a set of cascade linear SVMs that are learned corresponding to specific events and camera views.