R
RCAAP - RSS GERAL
Guest
Breve resumo:
Info Adicional:
Autor:
Clica para continuares a ler...
Piracy at sporting events broadcasts has increased dramatically in recent years, particularly in soccer. This dissertation presents a system whose main objective is to detect illegal broadcasts of soccer matches by detecting visual evidence. Several tools are presented, such as shot classification, logo detection, facial recognition and jersey color segmentation, in order to identify illegal visual content present in the game broadcast. Shot classification was used to classify the type of view present in the current frame, initially using DenseNet, but later Yolov5, with an accuracy of 96%. Logo detection was used to detect and classify visual elements present in the game, such as the scoreboard, competition logo and TV channel logo, using Yolov5, with an accuracy percentage of 91.43% for the competition logo, 65.25% for the SportTV channel, 62.38% for Eleven Sports and 100% for BTV and 86% for the scoreboard. Facial recognition was used to identify the game by recognizing the players' faces, using OpenCV for face detection and FaceNet512 and ResNet34 for feature extraction. Euclidean distance and cosine distance were used to compare the facial features. This system obtained an accuracy of 97%. The color segmentation of the players' jerseys was used to identify the teams present on the pitch through the color of the jersey, using MediaPipe, with an accuracy of 84.9% for segmentation and 70% for team identification. Finally, a discussion of the API used and the webservice created is presented. In conclusion, image classification, logo detection and facial recognition have a more important role in detecting piracy in soccer match broadcasts than jersey color segmentation.
Info Adicional:
Piracy at sporting events broadcasts has increased dramatically in recent years, particularly in soccer. This dissertation presents a system whose main objective is to detect illegal broadcasts of soccer matches by detecting visual evidence. Several tools are presented, such as shot classification, logo detection, facial recognition and jersey color segmentation, in order to identify illegal visual content present in the game broadcast. Shot classification was used to classify the type of view present in the current frame, initially using DenseNet, but later Yolov5, with an accuracy of 96%. Logo detection was used to detect and classify visual elements present in the game, such as the scoreboard, competition logo and TV channel logo, using Yolov5, with an accuracy percentage of 91.43% for the competition logo, 65.25% for the SportTV channel, 62.38% for Eleven Sports and 100% for BTV and 86% for the scoreboard. Facial recognition was used to identify the game by recognizing the players' faces, using OpenCV for face detection and FaceNet512 and ResNet34 for feature extraction. Euclidean distance and cosine distance were used to compare the facial features. This system obtained an accuracy of 97%. The color segmentation of the players' jerseys was used to identify the teams present on the pitch through the color of the jersey, using MediaPipe, with an accuracy of 84.9% for segmentation and 70% for team identification. Finally, a discussion of the API used and the webservice created is presented. In conclusion, image classification, logo detection and facial recognition have a more important role in detecting piracy in soccer match broadcasts than jersey color segmentation.
Autor:
Clica para continuares a ler...