Object Detection and Tracking for automated video analysis of padel matches and generation of game statistics
Develop a machine learning algorithm able to track players and balls in padel matches footage and to gnerate game statistics
Requirements
- M.Sc. in Machine Learning, Data Science, Computer Science, Mathematics, Telecommunications, or similar
- Knowledge of Python
- Software development skills
- Basic concepts of computer vision and image/video processing
- Familiarity with cv libraries (OpenCV) and machine learning libraries ( PyTorch, Lightning)
Description
The automatic analysis of sports videos has gained increasing attention in recent years, driven by its potential applications in performance analysis, tactical evaluation, player development, and fan engagement. In the context of padel, a rapidly growing racket sport worldwide, the ability to automatically extract meaningful statistics from match recordings offers valuable insights for both professional players and enthusiasts.
This thesis aims to develop a computer vision algorithm based on machine learning techniques capable of automatically analyzing video footage of padel matches. The system will focus on three main objectives:
(i) Player detection and tracking – identifying and following players throughout the match using object detection and tracking methods. (ii) Ball detection and trajectory tracking – recognizing and localizing the padel ball to reconstruct its motion and interactions. (iii) Shot recognition and classification – identifying the type of shots performed (e.g., forehand, backhand, volley, smash)
The research will begin with a comprehensive review of state-of-the-art methodologies in sports video analytics, including multi-object tracking, action recognition, and pose estimation. Selected approaches will be implemented and evaluated on a dataset of padel videos, either publicly available (retrieved after an extensive review) or collected internally.
Based on the literature review, the focus of this work will be the development of a methodology grounded on existing algorithms from the state of the art, adapted and fine-tuned for the specific use case of padel video analysis. The proposed system will integrate or modify known techniques in object detection, tracking, and action recognition to meet the challenges posed by the padel environment, such as small ball size, frequent occlusions, and the presence of glass walls.
Starting from the model predictions, the goal is to project the detected elements onto a 2D representation of the court, thereby enabling the extraction of spatial and temporal statistics. These projections will allow for the computation of meaningful performance metrics, including average player positioning, movement heatmaps, and shot direction analysis.
The final output will be a reproducible framework for automated padel match analytics, capable of transforming raw video input into interpretable, data-driven insights about gameplay. Such a system would provide a valuable foundation for further research in sports analytics, as well as practical applications in player performance evaluation, tactical training, and fan engagement tools.