智慧型居家安全攝影系統之研究

Abstract

在本論文中我們針對智慧型居家安全攝影系統有關應用提出三種系統：以條件性熵(entropy)做為特徵選擇依據的人形偵測系統；使用適應性機率模型(APM)為主的多用戶人臉辨識系統，以及使用平均位移法做人形追蹤系統。條件性熵使用於排序及選擇獨立成分分析(ICA) 特徵，且應用支持向量機 (SVM) 來對選擇後的特徵做人與非人上的分類。多用戶人臉辨識是以適應性機率模型為基礎，每個人臉模型皆使用主要成分分析(PCA) 對五張不同人臉方向影像擷取特徵所建立而成。在實際應用上，以串接式Adaboost方式偵測人臉並使用直方圖等化法對影像亮度做調整。再應用區域群聚法來精準地定位影像中人臉的位置。另外，我們提出的人臉辨識法可以線上註冊新的用戶和更新用戶資訊，以達到較快速、簡單且實用的使用者目標。最後一種系統為平均位移人形追蹤系統。我們提出的追蹤機制是針對監視區域中的人類移動做監控，並且利用HSV色彩資訊與人形特徵獨立成分做持續追蹤。此外，卡爾曼(Kalman)濾波器被應用在缺乏顏色資訊情況下來預測人類位置。 在最後，我們的實驗結果展示出我們所提出的方法在即時應用上有不錯的表現。

In this dissertation, we proposed three systems which are feature selection based on conditional entropy for human detection, multiclient face identification using APM (adaptive probabilistic model), and mean-shift human tracker, for intelligent home security camera system. The conditional entropy is used to sort and select the ICA (independent component analysis) features, and classify it as human or non-human by applying SVM (support vector machine). The multiclient face identification is based on APM model. Each face model is built by PCA (principal component analysis) extracted from five different head orientations. In practice, the face is detected by cascade Adaboost method. The effect of environment illumination in an image is adjusted by histogram equalization, and a region based clustering method is applied to accurately locate the face region in an image. In addition, our proposed face identification method can online register new clients and update the clients’ information to achieve faster, simple, and functional user goals. The last algorithm is mean-shift human tracker system. Our proposed tracking mechanism monitors the movement of human within a surveillance area and keeps tracking by using its color information in HSV domain and independent component of human features. Also, the Kalman filter is applied to predict location of human for the condition in lack of color information. Finally, our experimental results show that our proposed approaches can perform well for real-time application.