PKU@TRECVID2009: Single-actor and pair-activity event detection in surveillance video

In this paper, we describe our eSur system and experiments performed for the surveillance event detection task in TRECVID 2009. In the experiments, we addressed the detection problems of two categories of events: 1) single-actor events (e.g., PersonRuns and ElevatorNoEntry) that require only whole body modeling and no interaction with other persons, and 2) pair-activity events (e.g., PeopleMeet, PeopleSplitUp, Embrace) that need to explore the relationship between two active persons based on their motion information. Our contributions are three-folds. First, we designed effective strategies for background modeling, human detection and tracking. Second, we proposed an ensemble approach for both single-actor and pair-activity event analysis by fusing One-vs.-All SVM and rule-based classifier. Third, event merging and post-processing based on prior knowledge were applied to refine the system detection outputs, consequently reducing the false alarm in the event detection. We submitted three runs (i.e., p-eSur_1, p-eSur_2 and p-eSur_3), which were obtained by using different human detection and tracking modules. According to the TRECVid-ED formal evaluation, our prototype has yielded fairly promising results over TRECVid'09 dataset, with top Act.DCR of 1.023, 1.025, 1.02, and 0.334 for PeopleMeet, PeopleSplitUp, Embrace, and ElevatorNoEntry, respectively.