具擴展性可處理累積數據的Wi-Fi RSS指紋定位系統

Abstract

因應日益增加之高精確度室內空間定位需求，無論Wi-Fi RSS 指紋定位演算法乃至包括他種信號之各種場景分析定位法等，皆須面對場景資訊量龐大、難以由尋常終端待定裝置之有限計算資源來儲存與解算之挑戰。為克服此一問題，本研究提出一可伸縮、可重組化之空間定位系統，以軟體容器所建構之彈性計算後台為基礎，可因應海量指紋資料累積與待定物多寡之需求，透過網路機動調集的計算資源，而容器化、模組化的系統設計，易於增加多個包括Wi-Fi RSS與異種定位信號之指紋資料庫、多種並存之定位演算法、指紋蒐集與建模、以及具資料視覺化功能之使用者介面等各種容器，皆可透過參數化的方式調控與重組，便於研究者進行定位演算法之開發、組合與優化。 本系統原型主要以Python語言實作，使用Google Kubernetes (k8s) 作為軟體容器控管之主要框架，內建加權K-近鄰法與支援向量機兩種Wi-Fi RSS 指紋定位演算法模組，並已運行於三座具備雙64位元Intel Xeon處理器(共八核心)、64GB主記憶體、6TB存儲空間之Linux伺服器叢集上，可同時定位多個包括樹莓派(Raspberry Pi) 為基礎之待定裝置。基於本系統，本研究驗證了一新穎之W-Fi存取點選擇與過濾機制，以Wi-Fi 存取點之感知比率為篩選基準，有別於傳統以信號強度為篩選標準之做法。於本原型系統在台大明達館一15.7x9.4x2.9 m3測試場域試運轉期間，以系統內建之WKNN演算法在權重W為距離倒數、鄰近K值取5點的條件下，本AP選擇機制於ASR>=0.2時可濾除近78%之所有測得AP總數，大幅提升運算效率，而仍控制定位準度 (90%累積分布下維持3米等級) 之誤差變動不多於1.8%。

Attributed to the uprising demand of high-accuracy in-door positioning systems (IPS), localization based on 802-11 (Wi-Fi) received signal strength (RSS) fingerprinting and even other scene-analysis approaches will inevitably face the challenges of the massive context information to accumulate, store, and process, which is intrinsically overwhelming to the limited computing resource by a typical endpoint device under test (DUT). In order to address it, this study proposes a highly scalable and reconfigurable Wi-Fi fingerprinting IPS featuring an elastic computing backend based on software containers. Such a system achieves scalability by dynamically subscribing computing resources via containerization. Its containerizing and structuring system design also incorporates multiple strains of positioning algorithms including Weighted K-nearest neighbors (WKNN) and supporting vector machines (SVM), as well as various fingerprinting databases including heterogeneous ones, fingerprint gathering and modeling mechanisms, and a web-based User Interfaces (UI) with data visualization support for developers and administrators. All the modules can be easily reconfigured via altering system parameters from the UI, enabling future researchers to develop and explore different positioning algorithms and algorithm ensembles. The very first trial implementation of that system, primarily written in Python and utilizing Google Kubernetes (known as k8s) containerization framework, has been operational on a 3-node Linux cluster with dual 64-bit Intel Xeon processors (totally 8 computing cores), 64GB RAM, and 6TB massive storage space on each server node. It is capable of positioning multiple RaspberryPi-based DUTs simultaneously. Accordingly, this study also validated a novel Wi-Fi AP selection mechanism based upon the AP sensible ratio (ASR) instead of the conventional signal strength in previous RSS-based approaches. During the pilot run phase of that prototype in a 15.7x9.4x2.9 m3 closed testing site within the Ming-Da Building at National Taiwan University, the ASR-based mechanism effectively filtered out ~78% of all recorded APs in the built-in WKNN positioning module (with weight [W] is the reciprocal of the distance, and the number of neighbors [K] is set to 5), significantly improving the computing efficiency while retaining the average positioning accuracy (3-meter range within 90% cumulative distribution), by suppressing its deterioration to be as low as 1.8%.