The present paper advances a robust video fingerprinting method to be deployed within second screen synchronization applications. The novelty consists in tracking camera recorded video sequences based on their visual content. In this respect, the advanced system creates synergies between two architectural modules, a bag of visual words framework ensuring robust and scalable query localization in the reference database and an aggregation block providing an accurate synchronization of the video sequences as well as the system's decision. The system was tested on a reference database of 14 hours of video content and on a query dataset of 5 hours obtained by the StirMark geometric random bending and by live camera recording. The evaluation resulted into average false alarm rate of 1.3×10 5, probability of missed detection of 0.05 and F1 score equal to 0.96