基於 Transformer 用於物件狀態分析之關聯度計算模型

Abstract

籃球分析系統是現代籃球運動中不可或缺的工具，其中判斷持球者更是重要的任務之一。傳統的做法是先透過物件檢測取得球員與籃球的檢測框，再透過球員與籃球的幾何關係，例如計算球員與籃球的Intersection over Union，或計算球員與籃球的中心座標距離，來判斷球員是否持球。然而我們發現這樣的做法很容易發生誤判，肇因於籃球比賽中存在著複雜的球員重疊情況，幾何關係無法準確判斷球員是否持球，從而出現誤判的情況，這對於分析持球者的任務來說帶來了極大的挑戰。為了解決上述問題，我們提出了「基於Transformer用於物件狀態分析之關聯度計算模型」，藉由加入球員的骨架資訊做為動作特徵，透過self-attention的方法來學習球員與籃球之間彼此的關聯性。實驗結果顯示，我們的架構可以透少量的訓練資料，得到92.3%的持球準確率，這個結果超越了傳統演算法85.1%的持球準確率。最終，更在非訓練使用之測試資料集獲得95.4%的持球準確率。Basketball analysis systems are essential tools in modern basketball, where identifying the ball handler is one of the most critical tasks. The traditional method is to first obtain the bounding box of players and the basketball through object detection, and use the geometric relationship between players and the ball, such as calculating the Intersection over Union or the center coordinate distance between the player and the ball, to determine who is ball handler. However, we found that such an approach is prone to misjudgment, due to the complex overlap of players in the basketball game. The geometric relationship cannot accurately determine who is ball handler. This poses a great challenge to the task of analyzing ball handler.In order to solve the above problems, we proposed” A Transformer-based Relationship Computing Model for Object Status Analysis” by adding the player's skeleton information as the action feature to learn the relationship between players and the ball through the self-attention. Experimental results show that our method achieves an accuracy of ball handler up to 92.3% based on a smaller dataset, surpassing the 85.1% accuracy of traditional algorithms. In the end, it achieved 95.4% in test dataset which different from the training dataset.