This paper presents a framework for real-time video analytics, tested in the context of volume crime detection and prevention. The framework consists on an annotation data model, called the Video Content Description (VCD), and the Rule Manager, an event detection strategy based on expert rules. It has been devised to separate the analysis of a semantic layer, where VCD units are used and Rules executed, from the low-level image processing modules (e.g. background subtraction, detection-by-classification). This feature makes this approach to be easily deployable into existing platforms that possess specific optimized image processing modules and therefore make use of advanced semantic analysis in embedded platforms in real-time. Tests on a project-specific annotated dataset of videos demonstrate the capabilities of the proposed approach.