基於Kinect之障礙物偵測與導引系統於嵌入式平台上之實現

Abstract

根據內政部在民國101年的統計，台灣全盲與視障人口的總數大約有五萬人，而且隨著人口的高齡化，這數字不管是在台灣或是全世界都在不斷的攀升中。如果我們可以完成一個信賴度高、強健、低成本且可攜性高的輔助系統，這將會是一個很大的幫助。本研究使用微軟所開發的Kinect作為感測器來獲取深度資訊與彩色影像資訊，提出一個強健的障礙物偵測系統，透過深度資訊的分析與彩色影像的輔助，每個物件可以正確地被分割出來，並且經由一些物件特性抽取的機制我們可以辨別出每個物件是否為我們所定義的障礙物，計算出障礙物的距離、大小以及位置，經過模糊系統推論出一個安全的行走方向，最後透過語音的方式有效地提供給使用者安全的方向導引。 本研究實作於Devkit8500D雙核心嵌入式平台上，並且於複雜、動態與黑暗的場景中成功測試，障礙物偵測演算法平均可達到96%以上的偵測率。

According to Ministry of the Interior, the total number of blind and visually impaired is about fifty thousand, and these numbers are increasing with the aging population in Taiwan and around the world. If we can accomplish a reliable and robust vision solution for blind people with technology, at an affordable cost, then it will have a tremendous impact. This thesis presents a robust obstacle detection system based on depth and RGB information obtained from a Kinect sensor. With analysis of the depth map and RGB image, segmentation is adopted to distinguish different objects according to related depth information. Obstacles extraction mechanism is proposed to capture obstacles by various object proprieties revealing in the depth map. We aim to assist the visually-impaired in detecting obstacles and provide a guide for safety. The guide, on size, direction of the obstacles, and distance to the obstacles, is inferred by fuzzy rules involved with the above information. Our system is possible to alarm to users effectively according to the guide by voice. The proposed system was successfully implemented on the Devkit8500D embedded platform, and fully evaluated in a real unfamiliar environment. The experimental results show the accuracy ratio of 96% for variety of challenging scenarios such as thin obstacles, close obstacles, and hanging obstacles, as well as dynamic and dark environments.