以改良型航位推算法與粒子過濾演算法實現在2.5D環境中的室內定位

Abstract

最近，使用慣性感測器的行人航位推算法吸引相當的關注。 行人航位推算法的主要缺點是加速度計和陀螺儀容易產生累積誤差。 被稱為零速度更新法的技術嘗試去辨識腳掌停在地板上的時間點來校準移動速度。 然而，它需要傳感器安裝在腳的底部，因此會造成過大的振動（當使用者正在跑步），這有時會導致測量的速度和方向有很大的定位誤差。 本論文提出兩個能夠自動校準的行人航位推算法模型並且集成2.5D室內地圖資料以及粒子過濾器。 行人航位推算法模型分別用於走路和跑步，各稱為走路速度更新法與跑步速度更新法。 這些模型需要兩個／三個傳感器安裝在上半身／大腿／小腿，以期能在正確的時刻更新使用者的走路／跑步速度。 然後，我們考慮在一個多樓層建築採取這兩種模型的定位，並集成2.5D室內地圖資料。 粒子過濾器是用來過濾使用者的潛在位置。 我們更觀察到，如果使用者通過一些特殊區域（例如電梯或樓梯），慣性感測器會出現一些特殊的特徵。 因此，我們可以考慮這些特徵的當作地標校準定位結果。 我們已經開發了一個系統原型，並且進行廣泛的實驗。 結果表明當使用者走路時，走路速度更新法的軌跡更接近原來的形狀。 跑步速度更新法在使用者在跑步時的表現比其他方法更好。

Using inertial sensors for pedestrian dead-reckoning (PDR) has attracted considerable attention recently. PDR's main drawback is that accelerometers and gyroscopes are prone to accumulated errors. The Zero Velocity Update (ZUPT) technique tries to identify the moment when the sole is on the ground to calibrate the moving speed. However, it requires that the sensor is mounted near the bottom of the foot, which sometimes results large positioning errors due to the excessive vibration in the measurement of the velocity and orientation (e.q., when users are running). This thesis proposes two self-calibrating PDR models and integrates them with the Indoor 2.5-D Floor Plan model by the Particle Filters. The PDR models called Walking Velocity Update (WUPT) and Running Velocity Update (RUPT), are used for walking and running, respectively. These models require two/three sensors mounted on the upper body/upper leg/lower leg, which collaboratively calibrate the walking/running velocities of users at proper moments. Then, we consider the localization in a multi-floor building by taking these two models with 2.5D floor plan. The particle filters are used to respect the user's potential locations. We observe that the inertial sensors which are carried by users who pass through the special areas (e.g., elevators or stairways) have the identifiable signatures. Thus, we can consider these signatures as the landmarks to calibrate the location estimation. We have developed a prototype and conducted extensive experiments of our models. Results show that the trajectories of WUPT are closer to the original shape when the users are walking. The RUPT performs better than others when the users are running.