Soccer Field Registration Based on Geometric Constraint and Deep Learning Method

Pengjie Li, Jianwei Li, Shouxin Zong, Kaiyu Zhang

Abstract: Registering a soccer field image into the unified standard template model image can provide preconditions for semantic analysis of soccer videos. In order to complete the task of soccer field registration, the soccer field marker lines need to be detected, which is a challenging problem. In order to solve the problem, we propose a soccer field registration framework based on geometric constraint and deep learning method. We construct a multi-task learning network to realize marker lines detection, homography matrix calculation and soccer field registration. Firstly, the input image is preprocessed to remove the background and occlusion areas and extract the marker lines to obtain the edge map of the soccer field; then, we extract the features on the edge map and on the standard template model image to calculate the homography matrix. In this paper, a two-stage deep training network is proposed. The first stage mainly completes the soccer field marker lines detection and the initial calculation of the homography matrix. The second stage mainly completes the optimization of the homography matrix using geometric constraint, which can provide more accurate homography matrix calculation. We propose to integrate the geometric constraint of the marker lines into the multi-task learning network, in which structural loss is constructed to import prior information such as the shape and direction of the marker lines of the soccer field. We evaluate our method on the World Cup dataset to show its performance against the state-of-the-art methods.