A high semantic representation for abnormal events detection in crowded scenes

Many crowd abnormal motion detection methods in video surveillance have been proposed in resent years. However, most of them are based on low semantic features, such gray value, velocity and gradient. Usually, low semantic features contain weak discriminative information of the scene. In addition, these methods often ignore important information in time and space dimension. In this work, a high semantic representation is proposed. Slow feature analysis(SFA) is adopted to provide high semantic representation. Then, a random walk model, which takes into account the spatio-temporal information, is used to detect the abnormal motions in crowd. We conduct extensive experiments on two datasets to demonstrate the effectiveness of proposed method. Experimental results suggest that our method outperforms the state-of-the-art methods.