Detection of anomalies in surveillance scenarios using mixture models

In this paper we present a robust and simple method for the detection of anomalies in surveillance scenarios. We use a “bottom-up” approach that avoids any object tracking, making the system suitable for anomaly detection in crowds. A robust optical flow method is used for the extraction of accurate spatio-temporal motion information, which allows to get simple but discriminative descriptors that are employed to train a Gaussian mixture model. We evaluate our system in a publicly available dataset, concluding that our method outperforms similar anomaly detection approaches but with a simpler model and low-sized descriptors.