利用RGB-D Sensors進行人類動作的分析

Abstract

本研究利用視訊資料進行人類動作的分析，目的在於發展一套通用的人類動作分析技術可以應用在不同的領域上，如公共安全：機場、地鐵、體育館、購物中心等公共區域或是大樓中的自動化監控系統，偵測是否有人有異常行為（例如破壞公共區域）。或者居家照護系統：偵測家裡的孩童或老人是否有跌倒或爬到高處等危險的行為，若有危險行為發生便通報家屬及照護人員。 本研究利用微軟所開發的RGB-D Sensors（亦即Kinect）來擷取人體3D關節資訊，並計算關節夾角當作人類姿勢的特徵向量，由於這些特徵向量維度極高，因此我們利用流形學習（manifold learning）之等構映圖（Isomap）進行降維，並在低維度的等構映圖空間進行基本動作的切割與分群。接著將每一群的基本動作給定一個語意上的闡述並形成一個編碼書（codebook），最後此編碼書可以用來對測試者進行動作的辨識。

In this study, we develop a human action analysis technology from video data, which can be applied to many different fields, such as the public safety monitoring system in subways, shopping malls, other public area or buildings to detect abnormal behavior; the home care system: to detect the danger behavior of children or elders, such as falling at home. We use RGB-D Sensors (i.e. Kinect), developed by Microsoft, to retrieve the body joints 3D information and to calculate the joint angle as the feature vector of a human posture. Since the dimension of the feature vector is very high, we apply the Isomap algorithm, which is a manifold learning approach, to reduce the dimension. Then, each atomic action is segmented and clustered in the Isomap space, and all of the clusters from a codebook. Finally, we can use the codebook to recognize the tester’s actions.