Avaliação de extração de características espaciais para análise de vídeos de futebol.

Abstract

The popularization of football, as well as the advent of digital TV, has aroused interest in automatic techniques for summarizing football videos. In this context, the present work aims to investigate existing techniques and propose new techniques for the problems of Lawn Segmentation, Field Line Segmentation and Outlet Classification. A database of tables of various matches football was specially built for the experimental evaluation of the investigated techniques. Related works in the segmentation area of Gramado, in general, are based on the notion of dominant color and heuristics, but lack an objective experimental validation. In this work, a template matching strategy was proposed, combined with the technique of closest neighbors, for segmenting the lawn. A validation of the results, with the aid of ROC curves, demonstrated that the results obtained from the proposed method were superior to those obtained from a set of existing approaches, reaching 98.32% True Positive Rate (TPR) and 16.08% True Rate False Positives (FPR), in the best parameter setting. The results of the revised works on Segmentation of Field Lines do use of the Hough Transform, but omit relevant details or use very simple algorithms. A description of the parameters used in the Hough transform, as well as the development of improvements in the Regarding the pre and post-processing stages, they are contributions of this dissertation in the segment segmentation aspect. In the best configuration, the results of TPR and FPR were 77.30% and 6.23%, respectively, which are subject to improvement in future extensions, using the Hough Circular transform. For the Outlet Classification, an MLP (Multi Layer Perceptron) neural network was used, receiving several attributes derived from the previous steps, whereas related works did not consider relevant characteristics, such as such as the number of lines and the number of blobs. The validation of the results in this problem was also obtained by means of ROC curves, reaching as best results TPR of 99.60% and FPR of 0.16%.