基於主動外觀模型演算法之人臉方向估測

Abstract

本論文為開發更符合人性化的人機互動介面，不同於其他系統需要裝載繁複的感應器於頭頂上，本系統利用常見的網路攝影機(Webcam)並搭配具備影像辨識及偵測技術之人臉辨識系統，架設於頭部正前方即可自動偵測並判斷使用者頭部方向做為操控機器人轉向之依據，讓使用者擁有更直覺及便利的人機互動經驗。 本系統流程主要分為三個部分：偵測、追蹤以及辨識。偵測的部分，偵測出人臉的大約位置，加上人臉追蹤演算法，使得系統更為即時及穩定。人臉辨識方法上使用主動外觀模型演算法(Active Appearance Model)，此演算法包含了人臉最重要的資訊：形狀及紋理，結合這兩個資訊可辨識出不同人臉方向，再依此控制輪椅的轉向，達到簡易直覺性操作的目的。本論文在主動外觀模型(AAM)匹配的成功率到達95.625%，此外總共測試了82574張影像，包含不同場景與光線，而人臉方向估測則有95.345%以上的正確率。本論文最重要部分是將人臉轉向的功能整合至機器人上，並且能流暢的執行，達成不同應用的結果。

In this thesis, unlike most of the control systems use lots of complicated equipments and sensors to achieve the goal of face pose estimation, our system, just uses a USB webcam as our tool to fulfill the demand of user-friendly interfaces. Furthermore, our system, which makes good use of automatic face detection and face-poses recognition, provides an effortless and comfortable human machine interface for all users. This system could be mainly separated into three parts: detection, tracking, and recognition. In detection, face positions are approximately found by detection algorithm. Moreover, the tracking algorithm which not only ought to maintain a stable situation but also reduce the search range. In recognition, the AAM (Active Appearance Model) is applied. AAM contains two essential information, shape and texture, which can distinguish face poses into three directions. The final results would control the wheelchair-Robot toward different directions. The experimental results demonstrate that the proposed approach performs a successful AAM fitting ratio of 95.625%. Furthermore, correct face-pose estimation ratio of 95.345% with testing total 82574 images under different lighting conditions and backgrounds. The contribution of this thesis is that we successfully integrate the face pose estimation with wheelchair robot, and also accomplish different applications.