由個人移動軌跡紀錄分析交通模式與活動據點

Abstract

個人行蹤除了受到行程安排的影響外，也顯示了一個人某個面向的生活型態，例如常用的交通模式與常去的地點可以顯示一個人的習慣與偏好。而有些地點在一個人的生活模式中扮演較重要的角色；藉由較重要的活動據點可以了解一個人的生活模式。受惠於近來附有全球定位系統功能手機的普及，個人行蹤軌跡的取得與紀錄更加容易。從軌跡紀錄中可以進一步分析其交通行為與停留地點，並找出重要的活動據點。 在本論文中，我們設計了一個軌跡管理服務網頁，提供使用者將紀錄下來的軌跡上傳，並於網頁上標記、管理、與檢視。將軌跡分段並擷取特徵值後，我們應用條件隨機場模型（Conditional Random Fields）於交通模式辨認，並與支援向量機模型（Support Vector Machine）比較其預測準確度。實驗結果顯示，條件隨機場模型因為考慮了時間關係，辨識準確度較支援向量機模型高。 此外，我們使用 OPTICS 群聚演算法將使用者曾停留的位置群聚為活動據點，並比較十種可作為預測據點重要程度的特徵值，如造訪次數、停留時間等。在十種特徵值中，以造訪次數多寡與停留時間長短之排序最能將使用者認知中相對重要的據點優先辨識出來。

The whereabouts of a person not only indicates her schedule, but also reflects her lifestyle. The transportation taken and the places visited indicate the habit and preference of the user. With the growing popularity of commercial GPS loggers and GPS-enabled mobile phones, the positions of a person could be obtained and logged, and further analyzed to infer the transportation taken and places visited. Moreover, some places are more significant than others in one's daily life. These significant places shapes the life of the person. In this thesis, we created a prototype of a trajectory management service to annotate and visualize the trajectories. We adopted machine learning techniques to segment the trajectories and extract their features, and used supervised learning approach to train probabilistic models. We modeled the transportation mode learning problem as a sequence labeling problem using linear-chain conditional random fields (CRF). We compared the CRF model with support vector machines (SVM), and our results show that CRF outperforms SVM, when temporal relationship is considered. In addition, we adopted OPTICS clustering to find the places visited by the user. Results show that, among ten measures we used, visit frequency and stay duration predict the most significant places more accurately.