智慧型手機穿戴位置及移動狀態辨識

Abstract

透過手機上的感測元件、軟體使用狀況以及資料探勘技術，次世代的智慧型手機應有能力感測使用情境如使用者的手機配戴位置、移動狀態、喜好、情緒、活動、環境、身理資訊、所在位置、搭乘的交通工具及心跳等。透過這些使用者情境，智慧型手機可以提供新穎的服務如hTC blinkfeed 及 Yahoo Aviate 桌面。在眾多使用者情境研究中，我們先針對兩個使用者情境去做研究分別為移動模式(靜態和步行)及手機穿戴位置(手、手臂、胸部、腰部和大腿)。這兩項使用者情境我們稱為"穿戴行為"。在此篇論文中，我們提出了一個兩階段的框架來同時偵測手機穿戴位置及使用者移動狀態藉由手機的感測器。在推論穿戴行為時利用分類器演算法產生判斷模組，並測試多種分類器演算法如REPTree、SMO、Voting、Bagging以及Boosting，發現REPTree以及Boosting演算法表現極佳，在個人資料訓練時準確度可以高達95%。另外提出使用大眾資料訓練的方法來解決分類器訓練過程中會遇到的冷開始問題，但準確度較低只有65%，所以最後提出混合個人資料跟大眾資料的方法來提升準確度，發現在混合40%個人資料時就有不錯的成效，混到80%時為準確度最高。我們建議在建造判斷模組時，使用80%個人資料跟20%的大眾資料來當作訓練資料並使用REPtree或是Boosting演算法。

According the sensor on the smartphone, usage record of software and technology of data mining, next generation smartphones have the ability to sense user contexts such as user mobility, device wearing position, location, activity, emotion, health condition. Many apps utilize user contexts to provide innovative services, e.g., hTC blinkfeed and Yahoo Aviate windows. Two of the most important user contexts are mobility patterns (still and walk) and device wearing positions (hand, arm, chest, waist and thigh). We call these two user contexts “wearing behavior”. In this paper, we propose a framework to simultaneously recognize smartphone wearing position and user mobility by utilizing sensor data from smartphones. Using the technology of data mining to create the wearing behavior recognition model, and compare the accuracy of many recognition models which created by different algorithm. Then, we found good performances from the REPTree and Boosting algorithm, the accuracy of Boosting algorithm can be 95%. And the other, we solve the cold start problem using the general models, build by the other users data, but the accuracy of this model just had 65%. So, we merger the personal data and general data to improve the accuracy, and found that merger with 40% of personal data can had good performance and merge with 80% of personal data had the max performance. We propose using 80% personal data and 20% general data to be the training data and building the recognition model by REPTree or Boosting algorithm.