Anomaly detection in traffic surveillance with sparse topic model

Abstract

Most research on anomaly detection has focused on event that is different from its spatial-temporal neighboring events. It is still a significant challenge to detect anomalies that involve multiple normal events interacting in an unusual pattern. In this work, a novel unsupervised method based on sparse topic model was proposed to capture motion patterns and detect anomalies in traffic surveillance. scale-invariant feature transform (SIFT) flow was used to improve the dense trajectory in order to extract interest points and the corresponding descriptors with less interference. For the purpose of strengthening the relationship of interest points on the same trajectory, the fisher kernel method was applied to obtain the representation of trajectory which was quantized into visual word. Then the sparse topic model was proposed to explore the latent motion patterns and achieve a sparse representation for the video scene. Finally, two anomaly detection algorithms were compared based on video clip detection and visual word analysis respectively. Experiments were conducted on QMUL Junction dataset and AVSS dataset. The results demonstrated the superior efficiency of the proposed method.