Crowd behavior detection by statistical modeling of motion patterns

The governing behaviors of individuals in crowded places offer unique and difficult challenges, and limit the scope of conventional surveillance systems. In this paper, we investigate the crowd behaviors and localize the anomalies due to individual's abrupt dissipation. The novelty of the proposed approach can be described in three aspects. First, we introduce block-clips by sectioning the video segments into non-overlapping spatio-temporal patches to marginalize the arbitrarily complicated and dense flow field. Second, we treat the flow field in each block-clip as 2d distribution of samples and mixtures of Gaussian is used to parameterize it keeping generality of flow field intact. K-means algorithm is employed to initialize the mixture model and is followed by Expectation Maximization for optimization. These mixtures of Gaussian result in the distinct flow patterns precisely a sequence of dynamic patterns for each block-clip. Third, a bank of Conditional Random Field model is employed one for each block clip and is learned from the sequence of dynamic patterns and classifies each block-clip as normal and abnormal. We conduct experiment on two challenging benchmark crowd datasets PETS 2009 and University of Minnesota and results show that our method achieves higher recognition rates in detecting specific and overall crowd behaviors. In addition, the proposed approach shows dominating performance during the comparative analysis with similar approaches in crowd behavior detection.