非參數主題模型於具異常偵測之行為辨識技術的研究

Abstract

在這個簡單異質環境感測器的時代，行為辨識在各領域扮演重要的角色，例如辨識行為與偵測異常。但是，大多數的研究只針對其中一種，所以我們認為一個廣義的方法應該從挖掘數據的高層描述開始，這樣我們的方法就可以適用於各種模型，而有不同的用途。本論文中，我們提出非參數主題模型的架構進而從五個月的簡單異質環境感測器資料中萃取高層次的資料樣式，我們的方法可以被用在房間中並考量各種環境感測器資料。我們有兩種結果，在質性結果中，我們將資料樣式視覺化，讓人可以搭配一些照片驗證就能夠看出異常的活動，所以我們成功地在這些資料中找出曾經上過新聞的異常人士;在我們的量化分析中，結果顯示我們的方法可以有效的壓縮資料，讓空間節省超過 99% ，平均而言還可以達到最好的辨識率超過87%準確率，在偵測異常中，我們的方法可以有效地呈現高層次資料，讓支持向量機可以偵測出異常。總結來說，我們的方法可以有效呈現高層次資料樣式，且被用來解決真實世界的異常活動問題。

In the era of simple-heterogeneous-environmental(SHE) sensors, activity recognition plays an important role in many applications, such as recognizing normal activity and detecting abnormal activity. However, most research was conducted for either one but lost generality to both. We, therefore, start by developing higher-level representations for any qualified model to perform prediction as apposed to designing a specific model to fit certain data. In this thesis, we proposed a framework by non-parametric topic modeling to extract pattern as a higher-level representation from 5-month SHE sensor data. Our method can be used in a room with multiple sensors deployed and allow sensor fusion. In experiment, we have qualitative results and quantitative results. In our qualitative results, we generated visualized patterns for humans to detect abnormal activity. As a result, we successfully detect an intruder who was once reported on the news. In the quantitative evaluation, we demonstrate that the proposed framework can significantly compress sensor data over 99% and achieve best performance over 87% of accuracy on average in activity recognition. In anomaly detection, the proposed framework can effectively extract descriptive patterns for one-class SVM to detect abnormal activity. To summarize, we have shown the effective results of the proposed framework and use it to solve a real world problem, anomaly detection in a room.