智慧家庭環境中的情境感知

Abstract

隨著物聯網技術的蓬勃發展，智慧型裝置越來越貼近人們日常生活，像是智慧型手機、穿戴式裝置，以及平板電腦。其中，穿戴式裝置更是日益普及，如何利用穿戴式裝置去提升生活品質便成了一項重要的研究議題。在這篇論文中，我們使用穿戴式裝置偵測使用者的行為，使智慧家庭提供情境感知，判斷使用者的需求並提供不同的服務。 我們使用穿戴式裝置上的加速度器及陀螺儀去進行使用者動作的判斷，並且搭配Beacon提供的位置資訊精確地判斷使用者的行為。我們假設在一個客廳中，使用者可能有以下八種行為：看電視、閱讀報紙、與家人聊天、躺在沙發上睡午覺、聽音樂、做瑜珈、享受按摩，以及隨意地走動。 我們利用學習特徵值去建立行為的模型，並利用該模型去做比較，以判斷不同的行為，而特徵值來自於穿戴式裝置的動作資料及Beacon的位置資料。當使用者進行這些行為時，利用已建立之模型去進行比較，以判斷使用者的行為。針對不同的行為智慧家庭能夠給予不同的服務，像是調整燈光的顏色、亮度，或是控制電視、音樂。 我們使用決策樹與隱馬爾可夫模型去建立行為模型，相較於決策樹、k-平均演算法，許多相關研究使用隱馬爾可夫模型達到較高的精確度，因此我們決定使用隱馬爾可夫模型作為模型的演算法。首先，我們利用決策樹去取得隱馬爾可夫模型需要的觀察狀態，並且使用維特比演算法去解出隱馬爾可夫模型目前的狀態。我們使用Weka作為我們的探勘工具去建立我們的決策樹，為了建立我們的模型，我們必須預處理資料，產生特徵值，接著使用Weka產生決策樹，並且驗證模型，最後我們能夠利用這些模型精確地判斷使用者的行為。 我們的優勢在於低計算、硬體成本，位置資訊利用Beacon放在沙發、按摩椅及瑜珈墊上，相較於使用攝影機去判斷使用者位置，我們使用Beacon能夠降低計算成本；此外我們只使用一個穿戴式裝置去收集使用者行為資料以降低硬體成本。另外，我們針對單純使用決策樹及使用維特比演算法的精確度、利用不同運算平台的時間成本進行比較。

With proliferation of the IoT technologies, smart devices are closely connected to our life. Many of these devices such as smart phones, wearable devices and tablets greatly improve the quality of our life. In particular, wearable devices are becoming more and more ubiquitous. As a result, research on utilizing wearable devices to improve life quality has become important. In this thesis, we focus on the utilization of wearable devices in a Smart Home Environment. Our research objective is to provide situation awareness so that a Smart Home can respond to the need of a home owner based on the accurate detection of his/her activities. We use an accelerometer and a gyroscope on the wearable device for action detection. Such action detection combined with location information provided by Beacons would enable us to accurately detect a home owner’s activities. Our research assumes a living room environment in a home where the user may watch TV, read newspaper, chat with other family members, lie down on the sofa to take a nap, listen to music, do yoga, enjoy massage or just walk around. We detect different activities based on model comparison. First, the models of activities need to be constructed by learning the features of activity related data generated from the wearable device and Beacons. Then, when a user performs the activity, the newly generated data can be compared with the established models to identify the type of activity. Based on the different type of activity, the Smart Home environment will respond accordingly with different services, such as providing different color and intensity of the light, turning on/off TV and playing or stopping the music. Our method of model construction is based on Decision Tree and Hidden Markov Model (HMM). Most of related papers show that using HMM can reach a higher accuracy compared to Decision Tree and K-means. Therefore, we decide to use Hidden Markov Model (HMM) as our algorithm. Besides, we use Decision Tree to retrieve the observation data we need and Viterbi algorithm to find out current activities based on Hidden Markov Model. To build our Decision Tree models, we preprocess the data by calculating features. After that, we use Weka to generate our Decision Tree and verify the models with sample data. As a result, we can classify user’s activities by these models. To further classify user’s activities, we build Hidden Markov Model by experiment results from Decision Tree and solve the hidden state using Viterbi algorithm. Our advantage lies in low computation and hardware cost. First, location information is retrieved from three Beacons placed on the sofa, massage chair and yoga mat to reduce computation cost as compared to the use of a camera. Second, we only use one wearable device for raw data collection to reduce hardware cost. Third, we compare the accuracy between with and without Viterbi algorithm and the tradeoff between data processing on a smart phone and on a backend server.