An AI-Driven Approach to American Football Offense and Defense Detection

Success in American football is built on the skillful planning and execution of strategic plays. Instrumental in the strategic planning process is the annotation and analysis of video footage. Given the great deal of effort required to manually annotate such footage, we intend to develop a system that automates this process, effectively saving time and money for coaches and athletes. Several similar works have been done in the past, however there are few that focus on the differentiation of offense and defense-a crucial step for performing further analytics and the major focus of this paper. Therefore, we propose a network architecture that leverages player positions to train a multi-layer perceptron network to perform this task. We compared the proposed network with three other experimental architectures based on their performance on real game data. Results indicate that the proposed network architecture achieves up to 98.03% accuracy.