KTH人類動作資料集上以光流法辨識人類動作

Abstract

從一九七零年代以來，計算機視覺領域就開始研究人類的動作，人類動作的識別一直是很活躍的研究領域。近年來人類動作的識別在安全監視系統、老人與幼兒的居家照顧上有大量的應用，例如老人跌倒偵測，就是個很實際的應用。目前人類識別的研究方法大多是利用局部特徵及時空特徵做人類動作辨識，雖有不錯的辨識準確率，但在特徵偵測與擷取上十分複雜且費時，本論文擬提出一簡單的特徵擷取法，針對六種單一人類動作：拳擊、揮手、拍手、跑步、慢跑與走路，建立各項動作的訓練模型，再以訓練完成的動作模型，對未知的動作影片以支援向量機 (Support Vector Machine SVM)做動作的辨識。

Research in machine perception of human activities has started in the computer vision community since the 1970’s. Recently, recognizing human activities have a large number of applications in the safety surveillance system and care for the elderly and children system. Methods based on local features and spatio-temporal local features are now widely used in human activities recognition. Although they have good recognition accuracy, but they are very complex and time-consuming in feature detection and feature extraction. This thesis proposes a method to represent a human activity by simple features, six human actions are considered, they are boxing, hand waving, handclapping, running, jogging and walking. A training model for each human action is built, and the training model is used to recognize human action by SVM(Support Vector Machine).