基於適應性區域搜尋粒子濾波器之無線行動定位系統暨其整合應用於機器人導航及行人追蹤跟隨

Abstract

在本論文中提出一個以無線區域網路為基礎的室內定位系統，除了可供行動裝置使用者定位導引服務外，並整合應用於輪椅機器人在導航運送與人員陪伴跟隨的服務。本文首先探討利用無線區域網路中的訊號強度(received signal strength, RSS)，以內差法(interpolation)、曲線擬合法(curve fitting)及適應性網路模糊推論系統(adaptive neural fuzzy inference system)，建立起室內環境下之無線傳播衰減模型(radio propagation model)，利用該訊號模型可將接收訊號強度映射至相對應實際空間中的距離，透過擴展式卡爾曼濾波器(extended Kalman filter)估測出所在之位置。 另一方面，為降低機器人定位系統的成本以及增加實用上的普遍性，首先我們提出使用基於核函數密度估計(kernel density estimation)建立無線地圖模型(radio map model)以描述WLAN無線訊號RSS與位置的相關性，可有效解決非視線傳播(non-line-of-sight, NLOS)下對訊號預測效果；另外，為了解決低密度WLAN以及動態環境下RSS的可能的雜訊干擾，我們提出一套即時演算法選擇高強度且較穩定之無線訊號基地台(access point, AP)，可減低系統計算複雜度並提升定位穩定度；在位置估測方面，我們提出可自適應區域搜尋的粒子濾波器(adaptive local search particle filter, ALSPF)，當出現不準確的位置估測時可自動依據最近的行動範圍進行重新估測，增加定位精確度與可靠度。實驗中我們完整地比較了多種WLAN條件、AP密度、AP選擇以及無線地圖模型等狀況，以驗證所提出的定位系統的可行性。基於定位結果，我們運用模糊行為決策之導航系統提供自主機器人導航運送之服務。 由於目前服務型機器人大量運用於室內人群擁擠的場所，然而這些非預期的行人皆不會出現在地圖上，即便傳統建立地圖節點的方法能夠非常完善詳盡，但在動態及具有不確定性的環境下，許多節點實為多餘的節點，對於即時的路徑規劃是一種多餘的負擔。為了使整合無線定位系統的機器人導航更有效率，我們提出基於自動擷取地圖特徵節點的路徑規劃，其所擷取出的特徵點為地圖中為最精要的位置，如交岔路口、空曠區域的中心或路徑的末端，而免去其他多餘的節點來改善路線規劃的效能，且亦仍可滿足無線定位系統的精度需求；另外為了使導航的任務皆能順利達成，我們使用基於模糊推論的行為模式決策來處理高度動態的工作環境。搭配定位系統以及輪椅機器人，我們提出的自動導航可提供使用者傳呼叫車、位置運送等服務。 由於傳統輪椅容易造成照護人員在多人擁擠環境或是室外上下坡段較大的負擔，且不利於照護人員與病患面對面交談，本論文提出利用市售的微軟Kinect感測器，基於整合深度與彩色資訊之人員辨識追蹤與同步定位建圖系統，可用於複雜背景與移動場景下的即時三維物體追蹤，影像追蹤演算法以概似函數(likelihood)整合深度資訊描述向量，並利用最大後機率法則(maximum a posterior)找出最可能的跟隨目標，實驗中我們將演算法實現於輪椅機器人，在多人以及擁擠的場合都有相當穩定的追蹤效果。當目標人員遺失，機器人將自動前往最後追蹤的位置附近進行搜尋，加上整合前述之人員行動定位系統，可告知機器人目前目標人員的位置，當超出搜尋範圍則切換為自動導航前往。 綜合以上本研究所提出的系統與技術，不但可應用於室內的智慧型運輸系統，提供輪椅傳呼服務、運送服務以及陪伴服務，期望能輔助老人、小孩等行動弱勢族群能有更好的行動力與生活品質。

In this research, a WLAN based positioning system is proposed to provide the location service for the mobile users. In addition, we also investigate the radio based robot localization and navigation with the low-density WLAN, and the human tracking with wheelchair robots. Since the WLAN received signal strength (RSS) information is available in most wireless technologies, radio localization is being considered a feasible solution. Due to the unpredictability of WLAN radio propagation such as the interference, reflection and multipath effects, the adaptive neural fuzzy inference system based on the supervised self-learning strategy is proposed to estimate the physical distances from RSSI values. Then the extended Kalman filtering is used to perform the location estimation for mobile terminal user in the real indoor environment. To provide the reliable robot localization in practical usages in populated environments, we propose a new radio localization for mobile robots based on RSS in low-density WLANs. The observation likelihood model is accomplished using kernel density estimation (KDE) to characterize the dependence of location and RSS. Based on the measured RSS, the robot’s location is dynamically estimated using a proposed adaptive local search particle filter (ALSPF), which adopts the concept of covariance matrix adaptation for correcting the system states and updating the motion uncertainty. To deal with low sensor density in large-space environments, we present a strategy based on the strongest signal with minimum variance (SSMV) to choose a subset of detectable access points (APs) for enhancing robot localization and reducing the computational burden. The proposed approaches are verified by realistic WLAN with low sensor density and high speed robot motion to demonstrate the feasibility and suitability of the work. Experimental results indicate that the proposed ALSPF provides approximately 1 m error and significant improvements over particle filtering. For autonomous navigation, we design and implement the sensor based dynamic path planning and fuzzy based navigation for our mobile robot. The proposed path planning using the most essential graphic features from the given map builds the global and local path planning. The topology graph is constructed by using the equal-potential field methods, and then the global path planning provides the shortest path with a global view. In addition, the local path planning is used for dealing with the noisy sensor information acquired from uncertain environments. Moreover, we establish behavior based navigation with a fuzzy inference system to perform specific tasks based on results of the dynamic path planning. The feasibility and efficiency in various conditions such as impasse escape, moving in a narrow hallway, obstacle avoidance, wall following and goal seeking are examined in experimental results. In order to reduce the load of the caregiver or companion (care staff), we also present an approach to visual SLAM and human tracking for a wheelchair robot equipped with a Microsoft Kinect sensor which captures RGB and depth (RGB-D) images simultaneously. Based on the environmental SURF features, we present the natural landmark based SLAM using RGB-D data. Meanwhile, a depth clustering based human detection is proposed to extract human candidates. Accordantly, the target person tracking is achieved with an online learned RGB-D appearance model by integrating histogram orientation of gradient descriptor, color, depth, and position information from the target person. The experimental results show the effectiveness and feasibility in crowded environments. Consequently, the developed systems are expected to not only provide the compact, convenient and autonomous robotic vehicle but also improve mobility, nursing and caring service of the senior citizen and physical disabilities.