Abstract
Crowd scenes analysis is becoming one of the most active researches in computer vision. Panic behavior is a key indication of the occurrence of an abnormal event within the human crowd, and its detection helps preventing disastrous situations. The detection techniques reported in the literature analyze the temporal variation of either the motion magnitudes, the motion orientations, the crowd density or people interactions. However, all these features contribute to the characterization of a crowd behavior and ignoring one of them may lead to the degradation of the detection performances. In the present work, our contribution is threefold. First, a novel feature is proposed. It allows to simultaneously take into consideration all the aforementioned characteristics in order to analyze the human crowd. Second, a sparse representation is proposed and aims to facilitate the distinction between non-panic and panic situations. Third, data related to a panic behavior are considered as outliers with respect to non-panic related data and are statistically detected. The approach proposed in the present study has four major advantages. First, it does not depend on the crowd density level. Second, its detection performances outperform the state-of-the-art techniques for most of the videos. Third, it is not restricted to specific panic behaviors like escaping, gathering, dispersion and so on; it is applicable to any of the panic behaviors. Fourth, it is simple and easy to implement.