Borut Ogrinc (2009) A vision-based system for detecting and tracking of advertisement billboards. EngD thesis.
The subject of the thesis is an implementation of a -based system for detecting and tracking of advertisement billboards, which is a part of a billboard replacement system for sport events broadcasts. In the thesis we present previous research concerning billboard detection, among which we choose a method proposed by G. Medioni, G. Guy, H. Rom and A. François in the article ``Real-time billboard substitution in a video stream''. We improved the mentioned system with state of the art methods for detection and tracking, developed a prototype of the system and tested its performance. Billboard detection was performed by finding corresponding regions in video frames and reference billboard images. Process of billboard detection is divided into the following stages: detection of covariant regions in video frame and reference image, computing of region descriptors, matching the descriptors and estimating a homography given a set of matched regions. We studied and tested different affine covariant region detectors and region descriptors. Based on testing we decided to use Maximally Stable Extremal Region (MSER) detector and Shape Context descriptor. Billboard boundaries were set by robust estimation of a homography between a reference billboard image and billboard in video frame. For homography estimation we used DLT (direct linear transformation) with RANSAC (Random Sample Consensus) for added robustness in cases of falsely matched regions. Efficient Maximally Stable Extremal Region (MSER) Tracking was used for billboard tracking. We developed a prototype and tested its accuracy and speed. We tested the accuracy of billboard boundary estimation, proportion of correctly tracked regions and processor time spent on different tasks. The developed updated system prototype for billboard detection and tracking is a good basis for future studies on the field of billboard detection, tracking and substitution.
Actions (login required)