Service and End of Rally Detection in Badminton Videos

Video analysis of points won and lost in a badminton match has become a valuable tool for coaches and players to analyze their matches. The analysis requires someone to annotate manually the start and end of the rally for each point for the coaches and players to determine the badminton strokes employed during that rally. Hence, to eliminate manual annotation, we present a technique to detect the start of service and end of rally. It utilizes court details and poses to accurately identify the service action from the other badminton actions. The characteristics of service action are utilized to obtain the features and arranged in a 2D format for a convolutional neural network (CNN) to detect the service and non-service. Gradients are used to detect the flight of the shuttlecock till it eventually comes to rest which signifies the end of the rally. In computing the end of the rally, the algorithm considers the missing shuttlecock flight coordinates and when the shuttlecock is out of camera view. Two videos of 4 and 12 min comprising 12 and 23 consecutive points are used to determine the accuracy of the proposed algorithm. The experimental results show 88.6% accuracy for a total of 35 rally points.