ABSTRACT
Recently, there has been a surge of interest in the computer vision community on automated crowd scene analysis; as a result, crowd behavior detection and recognition are topics of many reported works. To understand crowd behavior, a model of crowd behavior needs to be trained using the information extracted from video sequences. Since there are only behavior labels as ground-truth information in most of the proposed crowd-based datasets, traditional approaches for human behavior recognition have relied only on patterns of low-level motion/appearance features to build their behavior models. Despite the huge amount of research on understanding crowd behavior in the visual surveillance community, the lack of publicly available realistic datasets for evaluating crowds' behavioral interactions means there is no fair common test bed for researchers to compare the strength of their methods against real scenarios.
