以新穎指紋辨識位置之群集配置基於支持向量機應用在室內定位上

Abstract

群聚演算法應用在室內定位上可以增進定位準確度和減少運算量。儘管如此，傳統的群聚演算法無法直接的處理。而這個問題是目前在結合監督式與非監督式學習時會遇到的主要議題。而這份研究主要提出了一個基於支持向量機下的新穎群集配置演算法簡稱SVM-C，而此演算法成功地解決了結合群集配置算法及室內定位時會發生的問題。SVM-C 主要以兩個級別間的距離為判斷依據，而非像傳統群集配置演算法是使用兩個平均值間的歐式距離作為判斷依據。這篇論文主要時 坐在現實的無線網路環境中，而實驗結果也呈現出在傳統的定位演算法及維度排序下，相較於K-means 與SVC，定位的精準度分別增加了19.61%和15.31％。而在不同的定位演算法及不同的維度排序下，SVM-C 皆占了優勢。

Clustering approaches have been used in location fingerprinting systems to improve positioning accuracy and reduce computational overhead. However, traditional methods can not use the data collected directly, and this problem is the main issue in combining the supervised and unsupervised learning. This study proposes a novel clustering algorithm based on SVM called SVM-C and it solves the problem about the clustering algorithm applying in the classification. The SVM-C approach focuses on the distance between the classes. It utilizes the margin between two canonical hyperplanes to cluster them instead of using the Euclidean distance between two average points. This thesis applies the proposed algorithms to realistic wireless local area networks. Experimental results demonstrate that the SVM-C outperforms the K-means and SVC reducing the mean localization error by 19.61\% and 15.31\% respectively under the traditional AP-selection schemes. The experiments based on different fingerprinting approaches and different AP-selection schemes also confirm the advantages of the proposed algorithms.