基於影像掃描之人體姿態識別系統

Abstract

近年來，電腦輔助之人類行為認知應用越來越廣泛，例如行為分析可對於異常事件(如跌倒)進行偵測與預警；姿勢分類和姿態識別則常見於人機互動，利用不同的姿勢或姿態來表示不同的命令或是指示；活動表示則用來代表人體之活動，同時也可以人體以卡通化之方式做一呈現。但是多數之研究常需使用昂貴之設備，或是要求被觀測者使用較不方便的穿戴式感應器，使得成本偏高或是方便性不足。為了改善上述情形且加上即時性及隱私性之考量，本論文利用電腦視覺的方法，透過監視攝影機來建立一個注重隱私且具有較少資料量及運算量的簡化人體模型（僅識別人體影像之突出部分）。本研究成果期望提供給姿勢分類或是行為分析之研究者做更細部動作之分析。

In recent year, computer-aided human behavior understanding becomes important in a variety of applications. For example, behavior analysis can assist the detection and warning of a falling event. Posture classification and posture estimation can be used to generate commands in human-machine interactions. Human activity representation helps animate human actions. Most research efforts use expensive devices or wearable sensors, which may increase the cost and making people uncomfortable. In this paper, we propose a surveillance system based on computer vision to establish human gesture representations which can preserve personal privacy efficiently in terms of data size and computation cost. To that end, an approach to automatic recognition of rough human posture, through identification of extruding parts of human body, is developed. The identification results can serve as a base for future research works in behavior and posture analysis which need to analyze more detailed human actions.