基於支持向量機之身體對稱性用於疾病姿態分析之建模與驗證

Abstract

傳統上需要研究人體連續運動在空間中的變化時，會在各關節上使用反光球方便攝影機捕捉位置。但若想將其使用在診斷上，此方法的前置準備過於費時，在醫療上時間人力成本損耗大，相對來說一天就只能診斷少數病人。本論文利用Kinect本身不需額外在受測者身上裝置設備的特性，設計並驗證一系列的特徵來表達受測者的左右側身協調性與對稱性。

本研究針對擁有單側偏癱或無力症狀之患者，設計出一系列能夠在受測者實驗時表達左右半身平衡性、對稱性之特徵群集。藉由這些特徵群集，從中利用特徵選取方法中著名的主成分分析(Principal Component Analysis)，挑選出最能夠代表資料集的前幾大特徵組合。藉由使用監督式學習方法支持向量機(Support Vector Machine)，獲得能夠將資料區隔成偏癱者/健康者、左偏癱/右偏癱的數個超平面，將這些超平面建構成一個複合式的分類器，最佳化之後即為我們想要的診斷模型。在與受測者病例交相對照後，我們藉由自己設計特徵建立的診斷模型驗證後獲得84.6%的分類精準度，擁有令人滿意的結果。

Traditionally we set multiple reflective balls on subject’ joints during the gait analysising experiment.But this method takes too much time on setting-up and cost too much as medical diagnosis system.This research solve the problem by using Kinect as free contact sensing system,designing several features to measure subjects’ symmetry and verifying them.

In this research, we design set of features to measure body symmetry and dynamic balance ability during the experiment.By using Principal Component Analysis,these features will be projected into new space,and we get several new components combined by the original set of features that include most information.For separating healthy and hemiparesis subject, left and right hemiparesis,we use Support Vector Machine to build several models and get hyperplanes for separating.Finally we use these hyperplanes to build a layered-model and achieve 84.6% accuracy.