Section: New Results

Anomaly detection in crowded scenes

Participants : Juan Manuel Perez Rua, Antoine Basset, Patrick Bouthemy.

We have defined an original motion-based method to detect and localize abnormal events in videos of crowded scenes. The algorithm relies on so-called labeled affine flows, involving both affine motion types and affine velocity vectors, and on view-based crowd motion classes. At every pixel the crowd motion class is inferred from the affine motion model selected among a set of candidate models estimated over a collection of windows. Then, the image is subdivided in blocks where local crowd motion class histograms weighted by the affine motion vector magnitudes are computed. They are block-wise compared to histograms of normal behaviors with a combined distance. More specifically, we introduce the so-called local outlier factor (LOF) to detect anomalous blocks. LOF is a local flexible measure of the relative density of data points in a feature space, here the space of crowd motion class histograms. By thresholding the LOF value, we can detect an abnormal event in a given block at a given time. Comparative experiments on several real datasets demonstrated that our method is competitive with methods relying on far more elaborated models and exploiting both appearance and motion, while yielding superior performance over motion-based anomaly detection methods.