利用骨架特徵對執行姿態上具時空變異之日常活動進行辨識

Abstract

近年來，動作辨識是影像視覺領域熱門的研究主題。為了使系統能夠以最貼近人類，以及最自然的方式來解讀精細且複雜的動作，我們透過電腦視覺來設計此動作辨識系統。雖然近年來也有許多關於活動辨識上的研究，但若要提高系統的實用性，則資料中群內相異性與群間相似性的問題便日趨重要。因此，在本篇論文中，我們的目標為建造出一個以電腦視覺為基礎的動作辨識系統，藉由深度影像追蹤其骨架資訊，藉由分析其動作特性，建立具有鑑別力的特徵描述模型。 為了達到此目的，我們將獲取的骨架關節位置，以一自身為中心的球狀座標系統重新描述，以達到無關視點的活動辨識，在空間上的執行變異問題，利用高斯模糊削弱關節位置的絕對性，建立姿態的特徵模型。在執行時間的變異上，我們結合動態時間校正演算法，來解決每個人執行速度快慢不一的問題。除了姿態上靜態的描述，動態的動作變化也在特徵描述模型中加以考慮，藉由記錄關節的角度變化，分析出各關節在不同的活動中隱含的資訊熵及物理意義，然後結合各關節自身的速度資訊，形成動態特徵描述模型。最後，結合姿態與動作特徵描述模型訓練出一支持向量機，並以兩組具代表性的公開動作數據集加以驗證，實驗結果顯示，本論文所提出來的動作辨識法能有效的處理群內相異性與群間相似性的問題，這將有助於未來實際應用於人機互動的介面上。

Human activity recognition has become one of the most important areas of research in computer vision. Without extra markers attached to human body, the interaction between machine and human can be natural in a vision-based recognition system. However, it still remains some challenges such as intra-class variability and inter-class similarity. In order to solve the problems, this thesis presents a novel skeleton-based activity recognition methodology with depth sensors. Due to the variation of execution styles and speed for different individuals, a Gaussian blur mask is applied to model the joint position variation while dynamic time warping (DTW) is employed to align the temporal sequences. Besides encoding the structure, the motion is also characterized. Through recording the joint angle trajectory, the entropy of each joint can be evaluated and then be combined with the projected velocity feature. Moreover, the support vector machine (SVM) is applied to acquire the classification results. In our implementation, two challenging public datasets are used to simulate real situations in our daily living. The experimental results show that the proposed approach is discriminative for human activity recognition and performs better than state-of-the-arts. This approach can benefit to several applications such as human-machine interaction (HMI).