具自我修正能力之室內定位

Abstract

位置感知服務(Location-Based Service)是近年來堪稱殺手級的行動計算應用， 因此對於位置追蹤的需求也隨之俱增；在室外，GPS 是目前最廣泛使用的定位方 法；在室內，則以樣本比對法(pattern-matching)的無線網路定位常被探討，且運用 WiFi 者最具實用性。樣本比對法是透過事先在取樣位置收集鄰近基地台的訊號強 度，建立訊號地圖(Radio Map)的資料庫；定位時，則將目前的鄰近基地台訊號與 訊號地圖相互比對，再以其訊號特徵的相似度估算目前位置。 然而，樣本比對法目前尚存在許多瓶頸：(1)訊號地圖受環境因素的影響而出 現不同的分佈狀態，若不修正將影響定位效率，而環境因素的影響卻是時變的 (time-variant)；(2)建立訊號地圖需要收集大量樣本資料庫，十分耗費人力。 在本計畫中，我們將提出一項創新的觀念，稱為具自我修正(self-content)能力 的定位方式。所謂self-content，有幾個層次：(i)具有自我修正radio map 之能力； (ii)減少(甚至不需要)事前的training 來建立訊號地圖；(iii)僅利用現行智慧型手機上 的sensors，無需外在輔助基地台(如WiFi Access Points)，即可由手機自行定位。我 們將跳脫傳統樣本比對法的精神，以手機上的感測器、相機及擴增實境技術來達到 自我修正、自我定位之目的；根據上述目標，我們將探討三個研究方向：  以 WiFi 基地台相互監測訊號強度，建立具有自我修正能力的訊號地圖： 我們將允許基地台間相互量測訊號強度，這稱之為基地台間的量測 (inter-beacon measurement)；以此基地台間的量測資訊我們將可以預估訊 號地圖的變化，更可重建即時的訊號地圖。基地台間的量測的功能是在現 行新一代的WiFi 與Zigbee 所允許的功能，據此我們將開發數項具有自 我修正、自我調適能力的樣本比對技術。  利用社群貢獻機制或智慧型機器人(iRobot)自動收集訊號地圖，免除事前 的training 工作：我們將提出一個由社群使用者協助收集訊號地圖的機 制，並依據貢獻資料者過去的可信度(credit)調整貢獻資料的權重，進而 達到隨時擴充訊號地圖之目標。此外，我們更將開發一個智慧型機器人， 可以週期性重建或調整訊號地圖，以減少人工收集的誤差和成本。我們將 整合智慧型機器人自動收集的訊號地圖及社群貢獻的訊號地圖，進一步達 成self-content 定位之目標，並設計方法來評價這些訓練資料以增進定位 精確度。  開發創新的無輔助訊號室內/室外定位技術：我們預期利用智慧型手機上 的相機、感測元件來標記或辨識室內/室外標的物，同時利用慣性元件中 的e-compass 來量測各標的物的方向角，最後透過三個以上的標的物的方 向角計算提供定位服務。這方法將跳脫過去需要外在AP 或基地台發射訊 號並收集訊號強度的依賴，達到手機自主定位之目的，目前尚未見其他團 隊有相似之成果。 上述目標均極具挑戰性，本特約人員計畫中，我們將利用三年時間逐一開發創新的 方法及建立雛型系統平台。我們預期設計出具自我修正能力、自動收集位置資料庫、 無需外在輔助訊號之自主定位等方法；此外，我們的方法將充份利用現行智慧型手機 上的相機、感測元件、通訊界面等，對外部infrastructure 的依賴性很低。我們相信在 self-content localization 議題上，可以成為國際一流的團隊。

Location-based services (LBSs) have been recognized as a killer application in the mobile computing and wireless communications fields. Therefore, the demand for better location-tracking and localization techniques also increases. In outdoor environments, GPS is the widest used localization technique so far. In indoor environments, pattern-matching localization techniques have been intensively discussed and most of the works have adopted WiFi networks as the infrastructure to serve as transmitters. In pattern-matching localization, we need a training phase to collect the RSS (receive signal strength) of each nearby WiFi access point (AP) at each training location. These RSSs are collected in a database called radio map. Then, during the on-line localization phase, a user device also collects nearby APs' RSSs and sends this characteristic vector to a location server. The location server then compares this characteristic vector against its pre-trained radio map. For example, the location with the most similar characteristic is retrieved as the device's current location. However, there are several bottlenecks in current pattern-matching techniques: (1) The radio maps usually change as the environments change. If we do not frequently calibrate the radio maps, the localization accuracy will be significantly impacted. However, the change of radio maps is always time-variant. (2) The collection of radio maps, which usually involves a large volume of training data, is very labor-intensive. In this proposal, we are going to develop a new concept, called self-content localization. There are several levels of challenges for a localization system to be self-content: (i) How can a localization system calibrate its radio maps automatically? (ii) How can we reduce, or even totally remove, the efforts in collecting radio maps in the training phase? (iii) Is it possible that a user device can determine its own location without using any auxiliary signal transmitted from any infrastructure network (such as WiFi network)? For example, a big challenge is: Can a smart phone use its own sensors and camera to calculate its location without relying on any external signal source. The goal of this project is to enhance the self-calibration capability of pattern-matching techniques and even gradually relieve the dependence on any infrastructure in typical pattern-matching techniques when conducting localization. This is what we mean by "self-content" localization. Based on these goals, we plan to investigate in three issues:  Self-adaptive radio maps: We will develop enabling techniques to allow APs to detect each other's RSS and use such information as indices to self-calibrate radio maps. First, we will switch an AP to the receive mode from time to time. Under the receive mode, an AP will be able to overhear nearby APs' RSSs. We call this inter-beacon measurement. An important observation is: If we conducted such inter-beacon measurement when collecting training data during the training phase, these measurements can be used as important indices of the environment factors when collecting our training data. Then, at the on-line localization phase, we can also ask APs to collect current inter-beacon measurements and use them as indices to select a "proper" radio map for comparison. As far as we know, both WiFi and ZigBee are able to support such capability. Based on this novel idea, we will develop several pattern-matching localization methods with self-training, self-adaptive, and self-calibrating capabilities.  Semi-automatic or fully automatic radio maps collection: We shall develop two ways to collect radio maps to reduce the collection overheads. The semi-automatic solution counts on community users or volunteers to contribute training data. However, volunteers are not always fully trustable. Therefore, we need a credit accounting mechanism to evaluate a volunteer's accountability. Also, the possibility of manual errors and even attackers needs to be taken care of. The fully automatic solution relies on robots to collect radio maps in a daily manner. We will use iRobot to collect radio maps. Our past work has shown how to program an iRobot. Note that an iRobot also needs to localize itself when going out and collecting training data. A novel RFID-based landmark solution will be proposed. Basically, iRobots will follow RFID tags sticking on the ground to trace itself. Then on its way roaming around, it collects training data. The fully automatic solution will completely eliminate manual data collection.  Self-content localization without auxiliary signals: The previous solutions all rely on some sort of training data. In this part, we will develop a self-content concept for localization. We plan to use the camera and sensors on a smart phone to realize the goal. The basic idea is to use the augmented reality (AR) concept to identify objects captured by a smart phone. In the meantime, the angle relative to the phone is calculated by the e-compass of the smart phone. By identifying at least three objects and calculating their angles relative to the smart phone, we will show how to compute the user's current location. Note that the above process does not rely on any auxiliary signal transmitted by external infrastructure. Therefore, this process is fully self-content. In addition, we will further extend our model to mobile cases, where the user may move around while identifying objects. The movement of the user will be computed by other sensors, such as accelerometer. Then, we will show how to conduct mobile self-content localization. All these ideas will be prototyped on existing smart phones. The above three goals are all very challenging. In this project, we plan to use three years to achieve these goals, through designs as well as prototyping systems. We will design location systems that can self-calibrate and self-adapt themselves to environment changes by semi-automatically or fully automatically collect training data. In addition, we will further develop self-content and even mobile self-content solutions via existing smart phones by utilizing the camera and inertial sensors on them. Such self-content localization solutions greatly improve over existing systems because no auxiliary signals and no infrastructure are needed. We believe that through the success of these developments, our team will become worldwide well-known team in the LBS society.