GloFANet：融合全域特徵以提升資料利用效率之足球場關鍵點偵測架構

Abstract

在現代足球賽事的影像分析中，足球場關鍵點偵測對於球員定位、戰術分析與進球事件判斷等應用具關鍵作用。傳統方法如尺度不變特徵轉換（SIFT）與霍夫直線轉換（Hough Line Transform）雖具尺度與旋轉不變性，然在視角變化與遮蔽情境下表現有限。深度學習技術雖改善此問題，但在樣本數量有限的情況下仍面臨挑戰。為解決此問題，本研究提出全域特徵增強網路（Global Feature Augmented Network, GloFANet），結合全域與區域特徵，引導模型聚焦於重要結構，以提升判別能力與穩健性。實驗採用 SoccerNet 之 25,148 張影像，搭配自行標註資料共 542,690 筆資料點進行訓練。結果顯示，GloFANet 在七類關鍵點中達成 85.86% 的平均準確率（mean Average Precision），較 2023 年 SoccerNet 冠軍方法提升 5.96%。特別在中線偵測任務中，僅 2,137 筆樣本即提升 20.48% 準確率，展現於資料受限情境下之高資料效率與應用潛力。Keypoint detection in soccer broadcasts is crucial for tasks such as player positioning and goal event detection. Traditional methods such as SIFT and the Hough Line Transform lack robustness to perspective changes and occlusion. Deep learning models, including High-Resolution Networks with multi-scale feature fusion, enhance performance but still face challenges in maintaining accuracy when training data is limited. We propose a novel method, the Global Feature Augmented Network (GloFANet), which integrates global and local features to help the model capture critical patterns and better recognize hard samples. GloFANet is trained on 25,148 SoccerNet images and our own annotations, totaling 542,690 data points. It achieves 85.86% mean Average Precision across seven keypoint types, outperforming the 2023 SoccerNet champion by 5.96%. In center line detection, with only 2,137 samples, it improves accuracy by 20.48% over the baseline. GloFANet remains accurate under limited data, demonstrating strong data efficiency.