Skip to Main Content
The detection of events that differ from what is considered normal is, arguably, the most important task for camera-based surveillance. Clearly, the definition of normal behavior differs from one application to another, and, therefore, approaches to its detection differ as well. In the case of intrusion monitoring, simple motion detection may be sufficient, such as based on background luminance/color modeling. However, in more complex scenarios, such as the detection of abandoned luggage, more advanced approaches have been developed, often relying on object path modeling. In this paper, we describe a new model for representing normality. Our model, that we call a behavior image, is low-dimensional and based on dynamics of luminance/color profiles, however it does not require explicit estimation of object paths. The process of estimating visual abnormality is then a simple comparison of training and observed behavior images, that we call behavior subtraction. We describe a new practical implementation of our model that is based on average activity. It is easy to program and requires little processing power and memory. Moreover, it is robust to motion detection errors, such as those resulting from parasitic background motion (e.g., heavy rain/snow, camera jitter). Most importantly, however, the method is not content-specific, and, therefore, is applicable to the monitoring of humans, cars or other objects in both uncluttered and highly-cluttered scenes. We support these claims by including various experimental results, from urban traffic, through sport scenes to natural environment.