Object-size invariant anomaly detection in video-surveillance

Nowadays, there is a growing demand for automated video-based surveillance systems due to increase security concerns. Anomaly detection is a popular application in this area where anomalous events of interest are defined as observed behavior that stands out from its context in space and time. In this paper, we present an approach for the detection of anomalous motion based on the extraction of object-size features that is independent of object size and video resolution. The proposed approach relies on a variable spatial window based on object size that has shown robustness in scenarios that present motion of objects of different sizes. We propose a system composed of four building blocks: background subtraction, feature extraction, event modeling and outlier detection. The proposed approach is evaluated on publicly available datasets which contain instances of abandoned objects of different sizes (considered as anomalies). The experiments carried out demonstrate that our approach outperforms the related state-of-the-art in the selected datasets. The proposal can identify anomalies associated to objects with different sizes and motion without increasing the number of false positives.