以多異質感測器進行日常生活行為連續辨識之研究

Abstract

日常生活行為辨識是用來達到主動服務以及自動監控的一項關鍵科技。我們希望能從一段包含未知行為的感測器資訊中，連續辨識出發生的行為與時間。在這個論文中，我們透過感測器序列的資訊，不斷的去判斷出每一個時間點發生的行為來達到連續辨識的目的。 透過混合多種異質的感測器有助於我們分辨出各種行為，然而，異質感測器在資訊呈現形式上往往有很大的差別。我們希望能發掘不同模型在整合這些資訊時的特性，在我們的研究中比較了不同的模型在這個問題上的適用度，包括隱藏馬可夫模型 (Hidden Markov Model)，%階層式隱藏式馬可夫模型(HHMM)，條件隨機場 (Conditional Random Field)，以及結構式支持向量機 (Structural Support Vector Machine)。 實驗結果說明，鑑別式模型如條件隨機場，以及結構式支持向量機對於整合感測器較為有效，其準確度明顯高於隱藏馬可夫模型。%其餘兩種模型。其中結構式支持向量機對於各種不同形式的感測器都能擁有相當好的結果。除此之外，我們引入了數種重疊特徵提取的方法，使用這些特徵值能夠進一步的改善準確度，在使用的這些特徵後，條件隨機場跟結構式支持向量機得到了相當接近的準確度。 為了提供主動的服務，我們比較了數種不同的即時辨識方法。在我們所比較的方法中，On-line Viterbi得到了最佳的單位時間準確度，然而卻會產生相當多不必要的服務。我們提出了Smooth On-line Viterbi方法來改善這種情形。

Recognition of daily activities is an enabling technology for active service providing and automatic in-home monitoring. In this thesis, we aim to recognize activities in a long sensor stream without knowing the boundary of activities. We formulate this continuous recognition problem as a sequence labeling problem. The activity is labeled every a fixed interval given the sensor readings. Fusing multiple heterogeneous sensors helps disambiguate different activities. However, these sensors are very diverse in readings. To evaluate the capability of models in dealing with such diverse sensors, we compare several state-of-the-art sequence labeling algorithms including hidden Markov model (HMM), linear-chain conditional random field (LCRF) and SVMhmm. The results show that the two discriminative models, LCRF and SVMhmm, significantly outperform HMM. SVM$^{hmm}$ show robustness in dealing with all sensors we used. By incorporating proper overlapping features, the accuracy can be further improved. In additions, CRF and SVMhmm perform comparably with these overlapping features. For active service providing, we evaluate various inference strategies for the on-line recognition problem. On-line Viterbi algorithm achieves highest frame accuracy but suffers from high insertion errors that may cause unexpected services. We propose smooth on-line Viterbi algorithm to solve this problem.