multi-personbehaviorclassication.PDFVIP

The BEHAVE video dataset: ground truthed video for multi-person behavior classification S. Blunsden, R. B. Fisher Institute of Perception Action and Behaviour, School of Informatics, University of Edinburgh scott.blunsden@jrc.it May 26, 2009 Abstract Although there is much research on behaviour recognition in time-varying video, there are few ground truthed datasets for assessing multi-person behavioral interactions. This short paper presents the BEHAVE project’s dataset, which has around 90,000 frames of humans identified by bounding boxes, with interacting groups classified into one of 6 different behaviors. An example of its use is also presented. 1 Introduction In the past 10 years, there has been an explosion of research into the analysis of video data, particularly aimed at the detection of ‘abnormal’ human behavior (the definition of abnormal is usually defined on a paper by paper basis). The state of the art in this research has reached the point where human targets can generally be reliably detected and tracked in all but extreme conditions (poor lighting, severe and sustained occlusion). With that success, research has been concentrating on analysis of individual behaviors [5]. What has not received as much research effort so far is recognising the behavior of groups of people. Some notable examples are European handball play classification [2], American football play classification [12], basketball play classification [17] and in a more general surveillance context by Hakeen and Shah [10]. The key to making progress in a problem are potential algorithms and publically available benchmark datasets for researchers to compare algorithms. There are several potentially useful algorithmic frameworks for group behavior classification, e.g. Hidden Markov Models, Coupled Hidden Markov Models [16] and Conditional Random Field models [3, 4]. In the case of video sequence analysis, ground-truthed video sequences are essential. Unfortunately, they are also very time-c