A Two-Stage Homography Matrix Prediction Approach for Trajectory Generation in Multi-Object Tracking on Sports Fields

Abstract

Homography estimation is a fundamental topic in computer vision, especially in scenarios that require perspective changes for intelligent analysis of sports fields, where it plays a crucial role. Existing methods predict the homography matrix either indirectly by evaluating the 4-key-point coordinate deviation in paired images with the same visual content or directly by fine-tuning the 8 degrees of freedom numerical values that define the matrix. However, these approaches often fail to effectively incorporate coordinate positional information and overlook optimal application scenarios, leading to significant accuracy bottlenecks, particularly for paired images with differing visual content. To address these issues, we propose an approach that integrates both methods in a staged manner, leveraging their respective advantages. In the first stage, positional information is embedded to enhance convolutional computations, replacing serial concatenation in traditional feature fusion with parallel concatenation, while using 4-key-point coordinate deviation to predict the macroscopic homography matrix. In the second stage, positional information is further integrated into the input images to refine the direct 8 degrees of freedom numerical predictions, improving matrix fine-tuning accuracy. Comparative experiments with state-of-the-art methods demonstrate that our approach achieves superior performance, yielding a root mean square error as low as 1.25 and an average corner errror as low as 14.1 in homography transformation of competitive sports image pairs.