Ball Tracking, Event Detection, Video Summarization of Football Vishwa Mangal Mentor: Dr. Aditya Nigam • Ball Tracking done using STRUCT (structured output for kernel Tracking). • Ground truth for the first few frames are given and then it tracks the ball in subsequent frames. • Stores the features from the ground truth and searches in radius of “R” which best match the feature. • Audio summarization using PRAAT language. • High Pitch implies the important moment of match. • The threshold was found at 60% of the mean pitch. Raw Data collection of Forehand, Backhand, and Service Prediction of player using YOLO in each frame. optical flow between the frames Using MBH to remove the noise. Train using 3D - CNN 3D - CNN Working Of YOLO YOU ONLY LOOK ONCE Data Specification • Australian open final 2017. • video to frame at 15FPS. • Manually classified frames into forehand, backhand, service. • 30 instances each of forehand, backhand, and service. Performance • Train/Test split at 50/10 for 2 class and 80/10 for three class. • On testing with two class i.e forehand and backhand, we got the 100% accuracy. • On testing with three class we got the accuracy of 70%. Future Work • Classify the strokes in untrimmed videos using T - CNN. 100% accuracy with 2 class. 70% accuracy with 3 class .