基於非對稱U-Net實現微小且快速移動之物體檢測網路

Abstract

本論文旨在探討物件偵測在微小、快速且特徵不明顯的物體上的應用。為了改進比賽戰術並提升技能，專業運動員和業餘玩家經常使用手機或相機記錄他們的練習和比賽。隨著這一領域的興起，越來越多的研究人員開始結合深度學習模型與運動分析，以提供更全面的見解。物件偵測是其中的關鍵任務，因為識別物體的位置可以提供有價值的資訊，如戰略分析。然而，針對如羽毛球這樣快速移動且模糊的物體進行追蹤的研究仍然有限。TrackNetv2方法基於VGG-16和U-Net，通過熱力圖檢測羽毛球的位置，但其架構需要大量計算資源，難以在實際應用中保持高效。為了解決這個問題，我們提出了一種名為TinySeeker的非對稱架構，這種新穎的架構不僅能精確的檢測羽毛球的位置，還能提高計算效率，在檢測精度和計算需求之間達到了最佳平衡，使其在現實應用中既實用又高效。實驗結果表明，Tinyseeker可以在保持精度的同時減少多達26%的計算量。這種架構在該領域標誌著一項重大進展，推動了物體檢測任務的可能性，並為未來的類似研究設立了新的基準。To refine strategies and augment skills, both professional athletes and amateur players routinely utilizecameras to document their practice sessions and games. As a result, an increasing number of researchers are exploring this field, aiming to offer comprehensive insights. Object detection is a pivotal task within this field, as identifying object locations can provide valuable insights, such as strategic analysis. However, only a limited number of studies have specifically focused on tracking fast-moving and indistinct objects such as a badminton shuttlecock. The preceding method, TrackNetv2, proposed the use of VGG-16 and U-Net, a heatmap-based approach, for badminton detection. However, the architecture of U-Net demands substantial computational resources in this paper. To tackle this issue, we present a pioneering asymmetric architecture named Tinyseeker inspired by U-Net. This novel model not only assures precise detection of the badminton shuttlecock's location, but it also champions computational efficiency. The reimagined structure strikes an optimal balance between detection accuracy and computational demands, making it a practical and effective solution for real-world applications. Experimental results show that Tinyseeker can reduce calculation up to 26% while remaining the precision. This architecture marks a significant advancement in the field, pushing the boundaries of what is possible within object detection tasks and setting a new benchmark for similar studies in the future.