A Deep Learning Framework for Soccer Object Detection and Event Recognition

Soccer analytics increasingly relies on automated systems for tactical insights. The proposed approach introduces a deep learning framework for object and event detection in soccer, addressing challenges such as occlusion and dataset variability. For object detection, two strategies are proposed using YOLO: a Single-Model Approach, which detects players, referees, goalkeepers, the ball, and other entities, achieving <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$93.7 \% \text{mAP}$</tex> on SoccerNet v2, and a Two-Model Approach that isolates ball detection to address size and motion challenges, resulting in 55% mAP on the same dataset. Event detection leverages VGG16 and EfficientNetB0, achieving maximum of 91.69 % accuracy on a seven-class dataset and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{7 1. 4 3 \%}$</tex> on SoccerNet v3 dataset. While ball detection exhibits dataset sensitivity-achieving <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$96 \% \text{mAP}$</tex> on RoboFlow but only 55% on SoccerNet v2-and event recall for free kicks and penalties remains limited, the framework offers scalable solutions for soccer analytics. Future work can include spatio-temporal modeling and adaptive architectures to enhance robustness across diverse match conditions.