Robust probabilistic logo detection in broadcast videos for audience measurement

In the last decades there has been a increasing interest in the development of computer vision strategies to automatically recognize logos in images/videos since several application contexts have arisen in which the logo detection task has also a huge economical relevance. In this paper a logo detection system is presented. The modular architecture consists of a decoder DVB-T/DVB-S, a workstation to drive/control/record and process videos and a console that enables the (even remote) management of various video streams coming from several decoders. The algorithmic pipeline runs on the workstation and it consists of a preliminary key-point detection based on speeded-up key point detection and a matching strategy based on an optimized version of RANSAC. The main advantages of the proposed solution are the capability to detect multiple occurrences in the same image and to keep detection accuracy even under heavy occlusions. Experiments were carried out on a publicly available dataset and on three challenging broadcast videos concerning a music show and two sport events (rugby and rally). The encouraging results make the proposed system a reliable measure of the visibility of logos, whose functionalities could be fully exploited if the cross-check with official audience rating will be carried out.