Perspective Transform Based YOLO With Weighted Intersect Fusion for Forecasting the Possession Sequence of the Live Football Game

Abstract

With billions of fans, football is a rapidly expanding sport that has proven essential to many nations and their citizens in particular. Some of the difficulties and the key challenges are analyzed in the football game which includes: pitch analysis, identifying the moments (pass, shot, assist and free kick), multiple cameras fixed across the stadium, dynamic background, ball localization in the entire ground, fast movements, mutual overlapping between the team members. The coarseness of analyzing the football events using traditional machine learning algorithm is relatively very less and the results of detection rate for various events are limited. A novel scheme is proposed for predicting the possession of the football game using spatio-temporal features and positional sequence. In this paper, we used Homographic Perspective Transform (HPT) for pitch segmentation and analysis. Also, a hybrid detector called You-Only-Look-Once (YOLO) with Weighted Intersect Fusion (WIF) is designed to acquire the inclusive context of moving objects (ball, players and referees) and they are monitored over the consecutive frames by analyzing their centroids using Geometric Midpoint (GM) approximation. Simple distance measure with thresholding technique is used to determine team's possession percentage. The experimental analysis shows the proposed system is efficiently identifying the positional sequence of football game events on both benchmark dataset (ISSIA) and the live matches between FC Barcelona Vs Real Madrid 2015 and Chelsea Vs Manchester City 2020.