基於粒子群最佳化演算法之無標記棒球投手三維動作追蹤系統

Abstract

利用多台攝影機實現運動員的三維動作追蹤一直是運動分析裡一個重要的部分。儘管市面上已經存在許多商業化的動作追蹤系統，但大部分不是費用高昂就是系統的建置十分複雜，這也導致了大部分的商業化動作追蹤系統都應用在實驗室的場景。相反的，大部分的棒球場都已經配備有多台的高速攝影機，如果能利用這些高速攝影機拍攝的影片來做三維動作追蹤與後續的運動分析，那對於運動產業來說會是一個很大的幫助。 本篇論文的目的，是利用從棒球場的高速攝影機拍攝的影片來做棒球投手的動作追蹤。首先，利用YOLO與AlphaPose做二維姿態的估測並得到二維姿態的資料。接著，利用兩兩相機之間的對極幾何關係重建出一個初步的三維骨架。然而這個初步的三維骨架會有許多錯誤的姿態與誤差。為了處理這些錯誤姿態與誤差，本篇論文設計了一系列前處理的步驟。 首先，三維資料會先經過一次過濾以去除一些明顯的誤差。接著，對有缺失的資料進行補值。最後，將整筆資料進行平滑化以便去除初步骨架中的微小抖動問題，此時便可以得到一個初步的結果。然而，平滑化的過程中也會導致在投球瞬間附近的資料被平滑掉，導致這些高速瞬間的預測值失準。因此，本篇論文利用粒子群最佳化演算法將這些被平滑掉的資料調整回到最合理的值。粒子群最佳化演算法非常適合應用在需要考慮多個觀測值的情境，因為粒子群最佳化演算法可以同時計算它的預測值對應到多個觀測值的誤差，並根據這個誤差再去做調整。然而，因為這個特性，粒子群最佳化演算法也可能會過度擬合到資料中的微小誤差。為了避免這個情形，在做粒子群最佳化之前，本篇論文設計了一系列前提來篩選掉不需要做最佳化的資料，並找到需要執行最佳化的投球瞬間。而為了解決遮蔽導致的姿態錯誤，本文設計了一個學習投手運動模型的機制。 最後，本篇論文利用從棒球場實際收到的數據來做實驗與分析，並對於本篇論文提出的方法進行測試與分析。

3D human motion tracking for athletes based on data captured by multiple cameras is an important part of sport analysis. Although there are some commercial human motion tracking systems, most of them are either high-cost or require complex setup, which makes such systems usually be used only in laboratory scenarios. On the contrary, almost every baseball field has high-speed cameras, it would be helpful if the data captured by these high-speed cameras can be utilized to do 3D human motion tracking. In the proposed system, the data captured by high-speed cameras at baseball field are utilized to track the 3D motion of baseball pitcher. For 2D pose estimation, the proposed system utilizes YOLO and AlphaPose to generate the 2D data from videos captured by high-speed cameras. After obtaining the 2D data, epipolar geometry is adopted to generate the initial 3D skeletons. However, there may be false detections and noise in the initial 3D skeletons. To deal with the false detections and noise, 3D data preprocessing is introduced in the proposed system. For 3D data preprocessing, a filtering process is adopted first to filter out obvious false detections. After that, the missing data are padded by interpolation. Finally, a smoothing process is adopted to remove noise. Since the smoothing process may also suppress the data around the pitching moment, particle swarm optimization (PSO) is adopted to deal with the error caused by the smoothing process. PSO is a useful tool to compensate the error, because it can adjust the value according to 2D data. To avoid overfitting, a series of conditions are set to filter out the data which has no need for PSO, and find the data around the pitching moment accurately. Furthermore, to deal with more serious false detections that cannot be filtered out, a motion model is designed and utilized. Experiments and analysis are provided to verify the proposed 3D human motion tracking system.