應用於機器人之基於影像人員活動偵測

Abstract

本論文之主要目的在研究以影像處理偵測人員之活動狀況，藉由單眼視覺攝影機擷取到的影像資訊，得以在環境中尋找人員的存在，並且完成五種人體姿態的辨識；在完成五種人體姿態後，藉由環境位置與姿態和停留時間的組合分析人員在家中常見的行為。本系統分為三個部份，分別為人員偵測、姿態辨識及活動偵測。人員偵測利用方向梯度直方圖(Histogram of Oriented Gradient, HOG)做為特徵，搭配支持向量機(Support Vector Machine, SVM)分類器來完成人員偵測。姿態辨識使用星狀骨架(Star Skeleton)做為特徵，搭配隱藏式馬可夫模型(Hidden Markov Model)訓練及辨識出站立、行走、蹲、坐及躺五種姿態。而在活動偵測設計中，本論文發展一套方法，利用環境物體特徵自動調整機器人移動後之環境邊界，得以偵測出人員目前所在的環境區域，再由人員在畫面中的位置與停留時間、姿態及偵測出人員所在的環境區域的組合，使用有限狀態機(Finite State Machine)辨識出在不同環境中可能的人員活動。經由實驗驗證人員偵測辨識率達95.33%，五種姿態之平均辨識率可達94.8%，另外，在不同的機器人視角下可以準確的辨識人在環境中的位置及其所對應的行為。

The main purpose of this thesis is to develop a vision-based human activity detection system employing a monocular camera. This system can be used for human-robot interaction in a home setting to provide service to people. A method is proposed to detect a human in acquired image frames. Five human poses are then recognized. The human activity detection system was designed by combining information from human location, human pose and the stay time. An environmental boundary detection method is proposed to determine the location of a human in the environment. This method uses features in the environment to automatically set environmental boundary, such that human location in the environment can be obtained. Satisfactory experimental results have been obtained with human detection rate of 95.33%. The pose recognition rate of five poses (standing, walking, sitting, squatting and lying) is 94.8%. Experimental study validates the performance of the developed method for human activity detection from different view angles.