基於優化無跡卡爾曼濾波器與迭代式LQR追蹤之無標記式單人三維人體動作估測系統

Abstract

近幾年來，三維人體動作估測一直都是個熱門的研究主題。從電影產業、復健治療到運動分析，越來越多的應用環境使得人們對三維人體動作估測的精準度與便利性有了更高的要求。隨者深度學習的興起，漸漸地有許多無標記式的估測方法被提出。但這些方法通常都會遇到缺乏室外三維標記資料的問題，使得提出的方法在現實情境中沒有辦法得到如預期的結果。 為了避開這個問題，一個僅基於二維人體關節偵測的方法在此篇論文中被提出。考量到直接將二維的偵測結果做三維重建可能會使得三維出測結果出現巨大的誤差，此三維重建結果還會經過骨架優化的步驟。此骨架優化由兩部分組成。第一部分為基於骨架模型的關節角度估計。第二部分則為動作平滑化。在關節角度估計中，除了關節角度的計算，來自三維重建的異常關節也會在被提出的異常分量排除無跡卡爾曼濾波器濾除以達到提升估計強韌性的目的。在動作平滑化中，除了位置之外，速度與加速度準確度這些高階次的指標也會在這一步得到顯著的提升。 最後，透過模擬與實驗，數據化的驗證所提出方法的效果與性能，以證明其可行性與精確度。

In recent years, 3D human motion estimation has been a popular research topic. From the film industry, rehabilitation therapy to sports analysis, more and more application environments make people require higher accuracy and convenience of 3D human motion estimation. With the rise of deep learning, many markless estimation methods have been proposed. However, those methods usually encounter the problem of a lack of outdoor labeled data so that the estimation results are not as good as expected in real-world situations. To avoid this problem, a method based only on 2D human keypoint detection is proposed in this thesis. Considering that direct 3D reconstruction of the 2D detection results may cause huge errors in the 3D estimation results, the 3D reconstruction results will also undergo the 3D skeleton modification process. The 3D skeleton modification process consists of two parts. The first part is joint state estimation based on the skeleton kinematic model. The second part is motion smoothing. In the joint state estimation, besides the calculation of the joint angle, the outlier keypoint from the 3D reconstruction will also be filtered out with the proposed outlier-component rejecting UKF (OCR-UKF) to improve the robustness of the estimation. In motion smoothing, in addition to position, higher-order metrics such as velocity and acceleration accuracy will also be significantly improved in this step. Finally, through simulation and experiment, the properties and performance of the proposed method are verified with data to prove its feasibility and accuracy.