基於高速球種定位系統之深度集成與漸進訓練策略

Abstract

近年來，將體育賽事與深度學習架構相結合在應用層面引起了廣泛地關注， 其中智能裁判、戰術規劃等結合神經網路進行輔助之需求日漸增加。在相關的應 用當中，深度學習架構通常扮演輔助角色，以幫助運動員或團隊針對比賽過程進 行分析，從而全面了解當前之比賽狀態，其中又以球類運動為甚。為執行高效之 戰術分析，在球類運動當中運動員和球種的位置檢測具有重要意義，其精確檢測 與否將大幅影響整體之戰術規劃結果。然而，部分球種之快速、體積小且不易預 測的特性使過往常見之物件偵測架構不易進行定位，成為了一個挑戰性的問題。 為此，於本文當中我們援引生成網路架構進行高速球種定位系統設計，並提出 ADEPTS 策略以針對該系統之訓練策略進行最佳化。ADEPTS 結合了多尺度特徵 融合和漸進式學習方法，使網路能夠更準確地捕捉高速球運動的軌跡特徵，同時 提高了訓練效率。研究結果表明，我們設計之高速球種定位系統可取得高定位精 度，且 ADEPTS 的加入可以額外減少該架構約 26.14%的訓練時間，這使其成為 實際應用中的實用和有效解決方案。In recent times, integrating sports events with deep learning architectures has attracted significant attention, resulting in an increasing demand for applications in this field. In regards to sports such as shuttlecock or tennis, the precise monitoring of player and ball positions holds significant importance. This task is indispensable for a thorough understanding of the game’s current state, serving both coaches and players. However, fast and unpredictable behavior makes extracting representative features a challenging issue that remains to be solved. To address this problem, we propose Advanced Deep Ensemble and Progressively Training Strategy (ADEPTS), which is an optimized training strategy designed for ball detection systems. ADEPTS combines multi-scale feature fusion and the progressive learning approach, allowing networks to capture the trajectory features of high-speed ball movement more accurately while improving training efficiency. Experimental results show that ADEPTS can significantly reduce training time by about 26.14% with high-resolution outputs. Additionally, it makes the network achieve even better localization accuracy, which makes it a practical and effective solution for real-world applications.