Image feature matching for scene extraction from single-camera videos

Abstract

Template matching is a widely-used technique for finding an image or a part of an image that matches the specified reference image, called a template. A tool was implemented using this technique to find template images in a soccer video. In the initial implementation of this tool, one approach of the template matching, template-based matching, was used but it lacked versatility and had some limitations for analyzing single-camera videos. In this thesis, we first analysed the existing template matching approach and identified its limitations. We then explored alternate approaches that could potentially address the identified limitations and found that feature-based template matching produced most optimal results for the tested videos. Feature-based template matching uses keypoints localization to detect the template in an image. The results showed that feature-based template matching could detect templates despite variances in the video quality and overcome some of the identified limitations of the existing approach.