PKU@TRECVID2010: Pair-Wise Event Detection in Surveillance Video

In this paper, we describe our system for the surveillance events detection task in TRECVid2010. We focused on pair-wise events (e.g., PeopleMeet, PeopleSplitUp, Embrace) that need to explore the relationship between two active persons. For our team had participated in the TRECVid SEDtask in 2009, we developed the system based on the old one. The improvements are three-fold. First, we refined the background subtractionmethod of last year. Some better background frames are automatically selected to train and update the background model and the background reconstruction is performed at pixel level instead of frame level. Second, we employed a MPL (Multi-Pose Learning) based method for head-shoulder detection, which can effectively improve the detection recall. Third, a structural SVM (SVM-HMM) classifier is employed for pair-wise events detection. According to the comparative results in the TRECVid SEDformal evaluation, our experimental results are promising.