基於RGB-D感測器之適應性人基於RGB-D感測器之適應性人員線上學習與追蹤及其在智慧型輪椅機器人的應用員線上學習與追蹤及其在智慧型輪椅機器人的應用

Abstract

在本論文中，我們利用Microsoft Kinect所提供的彩色與深度影像資訊為依據，並搭配快速線上學習目標人員(Online Boosting and Learning)的演算法，建立一套由三種分類器組合並串接一套線上學習外觀模型影像追蹤演算法的系統，讓機器人在不需事先建立任何資料庫或是額外感測配備輔助的情況下，可以短時間內快速學習並認識使用者，接下來再藉由系統不斷地線上自行更新目標人員相關訊息，及搭配影像追蹤演算法，來持續地追蹤我們所要跟隨之對象。除此之外，系統本身有辨識身分的能力，並且能同時追蹤多個目標人員並辨識其個別身分。本論文以實驗結果成功顯示我們的系統可以除了可以快速地找出我們要跟隨之對象外，由於系統本身不斷地更新目標物的資訊，因此在人潮擁擠、狹窄環境亦或是光線變化劇烈的情況下，均可以穩定地追蹤我們所要的目標人員並跟隨之。我們在實驗最後找了一位小朋友來當我們的對象，破除以往機器人所互動的對象都是系統開發者的使用限制，證明的系統可以適應不同使用者能力。 全球人口老化的現象，已成為各國重視的議題，由歐美國家經驗得知，65歲以上老年人口中，約有10% 的比例使用老人住宅，日本則在5%左右，經建會認為台灣地區目前僅有少數民間投資經營的老人住宅、養生村等，台灣人還是以「在地老化」或「在宅老化」為主流，因此居家照護需求很難下降。雖然已經有很多關於醫療輔助機器人方面的研究，然而在移動式服務機器人中，機器人與人之間的互動是很重要的一個議題，要達到機器人與人互動的功能，機器人必須要能夠適用於不同的對象，自行建立一套穩健的系統達到服務的目的。 另一方面，由於傳統輪椅使用上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，因此本論文提出利用市售的微軟Kinect感測器，整合深度與彩色的影像，建立起一套基於辨識及追蹤的系統，可用於複雜背景與移動場景下的即時三維物體追蹤，另外再搭配一套額外的影像追蹤演算法來提高系統的穩健性，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。 綜合以上本研究所提出的系統與技術，並且以“服務人群為目的”的觀念，使我們的系統可以適用於各種不同的使用者，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 system combined with multiple classifiers cascaded an RGB-D on-line tracking is proposed. Using the RGB and depth information provided by a Microsoft Kinect sensor, we build an on-line learning and training system based on three different boosting classifiers. The system gives the ability for wheelchair robot to cope with users without establishing any background knowledge or database as prior information. Moreover, the multi-classifier system is capable of calculating the confidence by itself during the whole updating process for continuous frames. Besides, cascaded with an on-line learned RGB-D appearance model assures that our system can deal with appearance changes, varying illumination or occlusion in a narrow or crowded environment. In addition, to examine the adaptation of the proposed system, we randomly select a child in a restaurant to be our tracking target. This not only breaks the stereotypes that the human-robot interaction should always be conducted by designers but also proves that our system can be adapt to any users. The world is in the midst of a unique and irreversible process of demographic transition that will result in older populations everywhere. As fertility rates decline, the population of aged peoples above 60 is expected to double from 2007 to 2050; in fact, the actual number will more than triple, reaching 2 billion by 2050. In most countries, the population of those aged peoples over 80 is likely to quadruple to nearly 400 million by then. Recognizing that a home is filled with memories and is more than just a place to stay, companies are engaged in accommodating the elderly for years of comfortable living. As the elderly age in place and their needs evolve, companies adapt services to meet the changes so that the homes remain well-kept and comfortable. There have been many technical advances in the realm of robotic research. But when it comes to mobile service robots, in particular their use for medical care, the idea of human-robot interaction becomes a very important issue. Robots need to establish a robust self-learned system, which can adapt to different users and achieve the purpose to provide service. Consequently, the developed systems are expected to not only provide the compact, convenient and autonomous robotic platform but also improve mobility, nursing and caring service of the senior citizen and those with physical disabilities.