Automatic Football Match Event Detection from the Scoreboard using a Single-Shot MultiBox Detector

During a football match, the information is manually collected by humans. However, the correctness of the football match data is difficult to check because of the game's speed, and thus, human errors can occur. This paper presents an automatic football match event detection from the scoreboard using a deep learning algorithm. The proposed method can reduce human error and performs the detection faster. In this study, the detection was trained with 30,000 data of Goals, Substitutions and Cards scoreboard from 68 matches of English Premier League 2017-2018 broadcast videos. The detection was tested with 80 sub-testing videos. These videos were prepared from 20 full matches broadcast videos, which consisted of 12 full matches from the year 2017-2018 and 8 full matches from the year 2018-2019. The proposed method contains three main steps: data gathering and augmentation, object detection for scoreboard visualization forms, and the event classification. The scoreboard detection is performed with a Single-Shot MultiBox Detector. The event classification employs the majority vote and time frame technique. The experimental results show an accuracy rate of 1.00 with the expected event scoreboards, comprised of Goal, Substitution, and Card events.