使用俯視深度攝影機之智慧家庭日常動作偵測系統

Abstract

在本論文中，我們提出一個新穎的室內動作偵測系統，用於自動記錄使用者日常生活的行為。我們使用俯視深度攝影機來實現動作偵測系統，這讓本系統比較沒有隱私問題且較不會對使用者的日常生活造成干擾。另外相較於傳統側向視角或是監視視角的彩色攝影機而言，我們的系統較不會受到遮蔽或是光線變化的影響。 動作偵測的目的在於指出一段影像串流當中的哪裡和哪一個時間點有我們在意的動作發生。在本論文中，我們將一個動作拆解成一連串依照特定順序排列的重要姿勢。我們使用隱藏變量支持向量機建立某個動作的模型，在模型當中我們學習了重要姿勢的外觀和重要姿勢在時間中應該如何排列。我們使用兩種互補的特徵來描述姿勢。第一個是深度差距直方圖，我們使用這一個特徵描述姿勢的外觀；第二個是位置標誌特徵，這一個特徵描述了人與地板和其他非地板的物體，譬如：床或椅子間的空間關係。 我們使用召回率─精準率曲線和平均精準度來衡量我們的動作偵測系統。實驗結果顯示我們系統的精準度和強韌性。

In this thesis, we propose a novel indoor daily activity detection system which can automatically keep the log of users’ daily life. The hardware setting here adopts top-view depth cameras which makes our system less privacy sensitive and annoying to the users. Moreover, in contrast with the traditional setting using side-view or surveillance-view RGB camera, our camera setting could avoid the problems of occlusion and illuminance change. The goal of action detection is to identify where and when the actions of interest happened in a video stream. In this work, we regard the series of images of an action as a set of key-poses in images of the interested user which are arranged in a certain temporal order. To model an action, we use the latent SVM framework to jointly learn the appearance of the key-poses and the temporal locations of the key-poses. We use two kinds of features which are complementary to each other to describe postures of human. The first kind is the histogram of depth difference value which can encode the shape of the human poses. The second kind is the location-signified feature which can capture the spatial relations among the person of interest, floor and other non-floor objects, e.g., bed or chairs. We use recall-precision curve and average precision (AP) to validate the proposed daily activity detection system and the experimental results show the accuracy and robustness of our system.