IMPROVEMENT OF ACTION ANNOTATION IN FOOTBALL MATCHES

With ever-increasing video consumption, video indexing is an essential tool that gives the ability to search through the video content. Action recognition in videos is not new but rather a highly researched field in computer vision. Within this field, recognizing actions in sports has bloomed over the years. Many approaches have been developed in recent years for various sports be it team sports or individual. Within this work, we propose an algorithm that automates the detection of actions in football matches on a frame-by-frame basis and classifies them into 40 predefined classes depending both on the angle and the length of action. After examining the broadcast video, we extract visual features from each frame to detect ball trajectory. Using this trajectory, we will group continuous ball movement into low-level actions using the clustering technique. These low-level actions are building blocks for high-level actions such as passes, kicks, dribble, etc. Results show that our approach works correctly 80 percent of the time. With such results, this approach can be the basis for other works such as next action prediction, and pattern recognition