動態環境下以主動式學習加強的自行重構之行為辨識

Abstract

本論文提出在智慧環境中學習及便是人類日常生活的問題，先前大部分的方法先收集人類行為的資料並學出其模型，然後再使用學得的模型來辨識人的行為，然而，人的行為習慣及環境的佈置可能會隨著時間而發生改變，造成行為的模式發生改變，這時舊有的辨識用的行為模型便過時了，使辨識率降低，必須要重新學習新的行為模型，但是重新收集學習用的行為資料並給予對應的行為標籤是件非常煩人且容易出錯的工作，在這樣的情況下，在更新行為模型時能降低人為的指導工作份量是件非常重要的事，本論文提出一個可以自我調整行為模型的行為辨識方法，它可以在動態的環境下同時辨識多種行為，並以較少的人為指導來跟著環境變動調整行為模型。

This thesis addresses the problem of learning and recognizing human daily activities in smart environment. Most approaches offline learn the activity model and recognize the activity in an online phase. However, the activity models can be outdated when the human behavior and environment deployment change. It is a tedious and error-prone job to recollect data for retraining the activity models. In such case, it is important to adapt the learnt activity models under one context to another context without too much supervision. In this thesis, we present a self-reconfigurable approach for activity recognition can reconfigure a previously learned activity model to infer multiple activities under a dynamic environment meanwhile requiring minimal human supervision for labeling training data.