Computer Vision Methods for 2D Ball Trajectory Extraction to Support Tennis Analytics

Trajectory estimation plays a crucial part in tennis analytics, as the ball’s flight path encapsulates a wealth of valuable information that can be exploited to enhance player’s performance. The goal of this thesis is to extract the complete trajectory of the tennis ball using suitable detection and tracking techniques, aiming to subsequently use this data, e.g., for evaluating shot effectiveness in future applications. To reach this goal several challenges need to be addressed, related to the ball’s physical characteristics – such as its small size and high speed – as well as external issues like occlusions from foreign objects, misdetection, or inactive balls in the background. Although previous studies have demonstrated robust and accurate trajectory extraction, they often rely on restrictive assumptions, such as the presence of a single ball. To overcome these limitations, the thesis proposes a vision-based system for detecting and tracking multiple objects using several cameras to filter and refine the ball’s true trajectory and enable its contextual understanding. To this aim, system functioning follows several phases. Beginning from the ball detection phase, different models are tested to predict the ball’s position, followed by the implementation of a multiple-object tracker developed to distinguish and separate the trajectories associated to each detected object. These operations are carried out using video footage of the same shot captured on-field from different perspectives. In this way, errors introduced during the detection phase could be identified and discarded by cross-verifying positions across different cameras. Preliminary experimental results demonstrated system’s robustness and accuracy in extracting ball trajectories under challenging conditions. Nevertheless, further refinement remains possible through future enhancements, such as integrating 3D projection techniques to define trajectories more precisely.