使用機器學習與分割取樣之人臉姿態追蹤方法

Abstract

人臉姿態追蹤一直是個重要的研究題目並且可以延伸出許多有趣的應用，其中，利用單一相機來追蹤姿態特別具挑戰性，這是因為在單一相機成像過程中失去了深度的資訊，本論文之目的在於提出一個強健且可達到系統即時性要求之追蹤演算法。 在本演算法中，我們使用影像平面之座標、目標物大小、平面上旋轉角度以及臉部左右轉動之角度共五個參數來描述目標物之狀態(state)，前四個參數是平面的資訊，而第五個參數則是立體的資訊，並且利用一個粒子濾波器(particle filter)來追蹤目標物狀態，由於前述特性，我們又將狀態空間拆為兩個部份，第一部份是平面狀態空間，而第二部份則對應至立體狀態空間，且將一般粒子濾波器之取樣分為兩個步驟，分割取樣(partitioned sampling)的好處是可以減少所需要粒子數目以符合即時性需求。 在平面狀態空間中我們使用的影像特徵包含顏色和輪廓，這兩個特徵都可以很快地從影像中萃取出來，而立體狀態空間我們則使用了相關向量機來估測一張包含人臉的影像對應到的臉部左右轉動角度。相關向量機(relevance vector machine)是一個機器學習(machine learning)的方法，好處在於訓練時間短且可得到一個很簡單的影像與臉部左右轉動角度對應模型，且此模型可去除掉表情改變以及臉部傾角改變的影響，且簡單模型有利於快速地估測角度。 最後我們更將所提供的演算法與一個主動式平台做結合，此主動式平台會跟隨目標物位置移動，以增加目標物被拍攝之範圍。

Tracking the orientation of human face has long been an important research topic which has many important applications. Tracking the orientation with a monocular camera is particularly challenging because the depth information is lost due to the perspective projection. This thesis aims to provide an algorithm to track orientation of a human face with efficiency and robustness. To solve this problem, we adopt the concept of partitioned sampling to decompose the state space with 5 dimensions, namely, translation, scaling, in-plane rotation and the yaw angle of the human face. In another words, the state space is decomposed into two portions, and one portion contains the parameters describing the planar motion of the target whereas the other contains the yaw parameter. The advantage of the state space decomposition is that we can avoid large amount of particles used for such state space and divide the efforts for the two portions with different sizes of the sample set. In this research, we first draw particles in the subspace of translation, scaling and in-plane rotation with simple cues such as color and contour. Then, we draw particles along the next subspace which contains only one dimension, the yaw angle of the target, and evaluate the yaw angle with the relevance vector machine (RVM). Here, RVM is trained for mapping an image patch containing human face to the yaw angle of human face. During the training process, we will add some perturbation of translation and scaling to the training samples of the yaw angle to make the prediction of face orientation robust to small translational errors. The learning based regression model is also insensitive to expression variation and unmodeled degree of freedom. Combining particle filter and RVM reduces the processing time and adds robustness to the performance of the system, thus making this algorithm applicable to human-machine interface with low-cost webcams and standard personal computers. The camera can be further mounted on an active platform so that the target to be tracked can be kept at the center of the image.