Automatic Camera Calibration on Football Broadcast Footage using AI

This thesis explores the integration of artificial intelligence (AI) in the sports industry, focusing on the outcomes of a work carried out in collaboration with Deltatre, a prominent sports and entertainment company, to enhance the spectator experience and improve analysis tasks. The work investigates the use of mobile broadcast footage for the calibration of cameras in optical recognition systems. By utilizing accurate metric information from easily accessible videos, new opportunities arise for analysis, processing, and visualization of dynamic sports events. Camera calibration (CC) is utilized in various services in the sports industry, such as graphics, officiating, video assistant referee VAR), and goal-line technology (GLT). However, the current approach relies on fixed cameras dedicated to each specific service and requires manual calibration on-site through a series of lengthy and complex operations. The goal here is to explore the possibility of introducing automation and leveraging AI models to streamline the process. After evaluating various solutions, this work adopts TVCalib, a recent camera calibration method. TVCalib employs a neural network consisting of two components: Segment Localization and Iterative Optimization. This approach demonstrated excellent results compared to existing literature. The first component utilizes the well-known UNet architecture, specifically deeplabv3, for accurate segmentation of field lines, crucial for subsequent calibration processes. The second component predicts camera parameters and lens distortion by minimizing a segment reprojection loss function. Through iterative optimization, TVCalib achieves highly accurate calibration results, enhancing the performance and reliability of optical recognition systems in sports applications. Evaluation based on metrics such as accuracy, compound score, and Intersection over Union (IoU), and comparison with state-of-the-art solutions consistently demonstrate improved outcomes. The integration of TVCalib within Deltatre's operations adds significant value to the company. Moreover, adapting the existing network to accommodate different camera angles and perspectives enables Deltatre to leverage its capabilities in additional contexts. The final product exhibits robust performance aligned with the company's specific needs, enhancing the accuracy and reliability of CC while opening doors to potential applications beyond the initial scope. This successful implementation showcases Deltatre's commitment to harnessing AI technologies and driving advancements in the sports industry. In conclusion, the integration of the TVCalib neural network within Deltatre's operations proves to be a valuable addition. This successful implementation enhances the overall capabilities of optical recognition systems and expands their usability across various contexts.