ETD

• machine learning
• neural networks
• sport analytics

This thesis is part of a broader project aimed at automatically extracting match statistics and player performance data from amateur padel matches using offline video analysis. To bridge the gap between a 2D video of a match and the 3D reality of the game, this work proposes a pipeline divided into three main tasks: court recognition, ball tracking, and 3D trajectory reconstruction.
First, to map the 3D world onto the image plane, a set of correspondences between 2D and 3D keypoints is required to solve the Perspective-n-Point (PnP) problem. A semantic segmentation model extracts key geometric features from the scene, and an analytical pipeline computes the keypoints' positions.
Second, the ball tracking module utilizes the TrackNet architecture to predict the coordinates of the ball from the video sequence. The pipeline also implements an outlier detection module and an event detection framework that analytically identifies events like bounces and hits.
Finally, the 3D trajectory reconstruction pipeline takes the 2D positions of the ball, the camera parameters, and the 3D initial and final coordinates of the flight sequence to fit a physically coherent trajectory. This process is performed by a Physics-Informed Neural Network (PINN), and its performance is evaluated against a standard Runge-Kutta (RK4) numerical integration method.