透過深度影像即時估計手臂動作之研究

Abstract

即時並且準確地估計上半身姿勢在人機互動的領域是一個重要研究，近年來的研究主要採用訓練機器的方式來即時辨識人體各個部位的資訊，然而，使用這種方法無法正確分析混亂的場景，例如不常見的姿勢和部分遮擋的情況，這是因為這類訓練方法如果沒有足夠的訓練資料就會導致錯誤的辨識結果，因此，本篇論文提出加入一個全新的視覺追蹤演算法來偵測手和肩膀的位置。使用追蹤的方法將會考慮到基於訓練資料的方法沒有參考到的時間資訊，當我們偵測完手部以及肩膀的位置後，我們利用計算區域覆蓋程度的方法來推算手肘位置並且獲取最終的上半身姿勢。我們的追蹤演算法結合了直方圖的計算、模版計算以及分群計算的結果，並且擁有以下優點：深度值或形狀變化的穩定性和較低的運算成本。實驗結果顯示，我們的方法可以即時辨識上半身姿勢並且比其他參考的方法獲得更好的結果。

Real-time upper body pose estimation is an important research topic in human-computer interaction. The recent state-of-the-art articles adopt a training-based method to acquire human parts in real time. However, this method still struggles to analyze cluttered scenes, such as unusual poses and partial occlusion. The reason is that in the training-based method, lack of training data will lead to improper results. In this thesis, we proposed embedding a tracking algorithm that combining a histogram-based response, template-based response and clustering-based response to localize human hands and shoulders. The tracking-based method utilizes temporal information that the related aforementioned method omitted. Our tracking method gains important advantages, including high robustness against depth changes, shape changes and low computational cost. After that, we utilized an area coverage comparison method to refine elbows and acquire the final estimated human upper pose. The experiments show that our method can obtain better performance than related methods in real time.