多姿態人臉辨識及其在機器人與人互動之應用

Abstract

本論文發展一套應用於機器人之多姿態人臉辨識系統。藉由Active Appearance Model(AAM)的方法找出人臉的形狀模型與紋理模型後，再對輸入的人臉資料利用Lucas-Kanade影像校正的改良演算法將形狀模型的特徵點以迭代的方式擷取到多姿態的人臉，並將擷取到的人臉影像資訊透過特徵空間的維度化減後，再進入倒傳遞類神經網路(BPNN)來得出受測者為哪一位家庭成員。整套系統已在實驗室開發的嵌入式數位訊號處理器(DSP)平台予以實現，並且也成功的應用在實驗室所研製的寵物機器人上。實驗結果發現，若以五位受測者做家庭成員多姿態人臉辨識，在UMIST資料庫與實驗室內建的人臉資料庫而言，其辨識率分別可達91%與95.56%，驗證人臉在不同姿態的情況下，所提出之方法能有效辨識出受測者的人臉。

In this thesis, a pose-variant face recognition system has been developed for human-robot interaction. In order to extract the facial feature points from different poses, active appearance model (AAM) is employed to find the position of feature point. The improved Lucas-Kanade algorithm is used to solve the image alignment. After obtaining the location of feature points, the eigenspace of texture model is reduced the dimension and sent to the back propagation neural network (BPNN). By using the BPNN, the proposed recognizes that which family-member is the user. The proposed pose-variant face recognition system has been implemented on an embedded image system of a pet robot. In order to test our method, UMIST and self-built database are both used to evaluate the performance of the proposed algorithm. Experimental results show that the average recognition rate of the UMIST database and self-built database in our lab are 91% and 95.56% respectively. The proposed pose-variant face recognition system is suitable for applying to human-robot interaction.