Automatic Statistics Extraction for Amateur Soccer Videos

Amateur soccer statistics have interesting applications such as providing insights to improve team performance, individual coaching, monitoring team progress and personal or team entertainment. Professional soccer statistics are extracted with labor intensive expensive manual effort which is not realistic for amateur matches. In this paper we develop a solution that automatically extracts action-related soccer statistics from a static camera pointed at the pitch. We implement a solution to player localization and action classification problem in human action recognition. Our method does not rely on player tracking, sliding windows, super voxels or construction of multiple hypotheses. Our work is developed with actual application in mind and a fully functional recognition pipeline is implemented, specifically tailored to meet the inherent challenges of action-rich soccer video.