Real-time video processing is steadily becoming a standard matter in a wide diversity of applications, including computer vision, surveillance, social networks, and many other. Unsupervised video interpretation is required to avoid the operative expense and faultiness of human-in-the-loop interpretation. However, robust, general purpose real-time unsupervised video interpretation and analysis appears to be among the most difficult processing tasks. In this work we present advances in real-time video stream processing, in particular in automatic key-frames detection. This detection procedure is based on feature analysis and evaluation of the features detected by the popular SURF algorithm. We present the implementation details, some interesting experimental results are shown, and we discuss some of the applications that our detection algorithm may have.Sociedad Argentina de Informática e Investigación Operativa