Temporal and spatial anomaly detection using Phase Spectrum of Quaternion Fourier Transform

In this paper, a novel method based on Phase Spectrum of Quaternion Fourier Transform (PQFT) for anomaly detection in crowd is presented. The proposed method mainly consists of two parts: quaternion feature representation and construct of abnormal saliency map. In the first stage, motion and appearance information are both considered and a quaternion feature representation is proposed for temporal and spatial anomaly detection simultaneously. In the second stage, QFT is first used for quaternion frames, and then normalize the amplitude spectrum. After that, the final abnormal saliency map is reconstructed with inverse QFT. Experiments on benchmark have proved that the proposed method detects abnormal events effectively and robustly and meets the real-time requirements.