排球比賽轉播影像之場地自動分割與幾何校正研究

Abstract

隨著運動領域對科學化的重視日益提升，如何兼顧客觀性與有效率地解析比賽過程成為當前的重要課題。現行排球資訊紀錄多仰賴人工或半自動系統，不僅耗時費力，且容易受主觀因素干擾。近年來，隨著轉播技術的進步，賽事影像資料的數量與品質持續提升，傳統的記錄與分析方式已難以因應日益龐大的資料需求，促使自動化分析技術逐漸受到關注。場地校正作為排球自動化資訊分析的核心基礎，對戰術分析與即時判斷具有重要影響。本研究運用電腦視覺技術，提出無需標註資料的場地自動分割與幾何校正方法，結合機器學習與幾何特徵分析，能自動從動態視角的影像中擷取關鍵場地區域，並利用高斯混合模型進行影像分群以辨識場地範圍，結合線段分類模型，進一步透過單應性矩陣完成幾何校正，將場地資訊映射至標準二維平面座標系統。本研究共收集21部國際排球賽事轉播影像，涵蓋四種不同場館，並於不同光照與視角條件下評估方法的穩健性與準確性。實驗結果顯示，影像分割精度達98.91%，場地校正整體IoU達98.14%。此外，本研究亦建立並公開涵蓋6837張排球轉播畫面的資料集，其中6077張完成影像分割及幾何校正的手動標註，為排球場地偵測技術的標準化與可重現性提供了參考資源。

As the field of sports increasingly emphasizes scientific approaches, achieving both objectivity and efficiency in match analysis has become a key challenge. Current volleyball information recording methods largely rely on manual or semi-automated systems, which are time-consuming, labor-intensive, and susceptible to subjective biases. With advancements in broadcast technology, the volume and quality of match video data have continued to grow, rendering traditional recording and analysis methods inadequate for meeting the expanding data demands and highlighting the need for automated analysis technologies.Court calibration is a fundamental aspect of automated volleyball information analysis, significantly impacting tactical analysis and real-time decision-making. This study employs computer vision techniques to propose an annotation-free method for automated court segmentation and geometric calibration. By integrating machine learning and geometric feature analysis, the method can automatically extract key court regions from dynamic camera views. A Gaussian Mixture Model is used for image clustering to identify the court area, while a line segment classification model, combined with a homography matrix, completes geometric calibration, mapping court information to a standard two-dimensional coordinate system.A total of 21 international volleyball match broadcast videos were collected, covering four different venues and evaluated under various lighting and viewing conditions to assess the robustness and accuracy of the method. Experimental results show an image segmentation accuracy of 98.91% and an overall court calibration IoU of 98.14%. Additionally, this study establishes and publicly releases a dataset comprising 6,837 volleyball broadcast images, of which 6,077 have been manually annotated for image segmentation and geometric calibration, providing a valuable resource for the standardization and reproducibility of volleyball court detection technologies.