PKU-NEC @TRECVID2012 SED - Uneven-Sequence Based Event Detection in Surveillance Video.
Ziwei Xia,Xiaoyu Fang,Yaowei Wang,Wei Zeng,Hongming Zhang,Yonghong Tian
2012-01-01
Abstract:In this paper, we describe our system for interactive and retrospective surveillance event detection task in TRECVid 2012. We focus on pair-wise events (e.g., PeopleMeet, PeopleSplitUp, Embrace) that need to explore the relationship between two active persons, and action-like events (e.g. ObjectPut, CellToEar, PersonRuns and Pointing) that need to find the happenings of a person's action. Our team had participated in the TRECVid SED task from 2009 to 2011. This year the new improvements of our system are three-fold. First, we introduce geometric constraints into the human detection. Second, we propose a robust tracking-by-detection approach with an optimized observation mode to address ID switching and tracking drifting. Third, an uneven-sequence classifier is employed for action-like event detection. Overall, we have submitted three versions of interactive results, which are obtained by using different human detection, tracking and events detection modules, and one version of retrospective result. For the definition of interactive task is ambiguous, we mainly introduce our work on retrospective task in this paper. According to the results in the TRECVid SED formal evaluation, our experimental results of retrospective task are promising.