Abstract
A two-stage framework for crowd anomaly detection in single-scene or scene-dependent surveillance videos is proposed in this article. The first stage generates several hypotheses corresponding to potential anomalous regions in a video frame and the second stage verifies them to reduce false alarms and identifies crowd anomalies. In the hypotheses generation stage, spatial and temporal derivatives are computed for each video frame and a saliency detector employing Hypercomplex Fourier Transform (HFT) is used to generate a saliency map. A threshold is applied to the saliency map to generate potential anomalous regions in the form of connected components. For each connected component, a set of 4 statistical features are computed and fed to the second stage which employs a Gaussian Mixture Model (GMM) as a verification method to yield the final crowd anomalies in the frame. The effectiveness of the proposed framework has been shown through results obtained on the UCSD anomaly detection benchmark dataset which contains two subsets namely Ped1 and Ped2 with a total of 48 test videos (9210 frames). Both frame-level and pixel-level anomaly detection results are provided using the widely recognized evaluation criterion in the domain and compared with the state-of-the-art methods. The experimental results show that the proposed framework obtains comparable results against the state-of-the-art methods.