從環境中的資訊辨識出使用者當下的行為是挑戰的,為了提供使用者更合適的服務,了解使用者的行為將能更正確地提供使用者所需要的服務。在過去的研究裡,為了簡化問題的複雜度,往往有家中僅有單一居住者的假設。本文的目標是建立一個系統能學習多位居住者的行為模式,並能即時的辨識家中各個居住者的行為。 本文的貢獻主要有以下三點:第一,茲因在非侵入性環境中資訊的不充足,多人行為辨識的資料鏈結並不容易被分類。於是我們嘗試設計能夠推論資料鏈結的模型;第二,由於每個居住者的行為有可能被其他居住者影響,人與人之間的互動必須被納入考慮,這樣的動機使我們提出了一個多人行為辨識模型;第三,手動標記資料鏈結是一個十分耗費人力的工作,特別是在多人環境中,因此我們實作了一個自動標記方法。
Reliable recognition of activities from cluttered sensory data is challenging and important for a smart home to provide more desirable services. Traditionally, most of prior works often assume that there is always only one resident at home for the purpose of simplification of solving a complicated problem regarding multiple-resident activity recognition. Therefore, the goal of this thesis is to build a system which learns multiple-resident activity models to facilitate reliable activity recognition of each resident. The main contribution of this thesis is three-fold. Firstly, due to the insufficiency of information in an environment using pervasive non-obtrusive sensors, data association for multiple-resident activity recognition is hard to be identified. Therefore, we aim to design a model which can infer the data association. Secondly, interactions among residents should be modeled since the activity of one resident may be influenced by another, and this concern motivates us to propose a multiple-resident activity recognition system which can adapt to the interaction between residents. Thirdly, because manually annotating data association is laborious, especially in an environment involving multiple residents, we design a mechanism which can automatically annotate the data association.