頭載多影像與深度感測器做盲人室內行走之自動導引

Abstract

本論文利用頭戴多台影像與深度感測器來實做盲人可以在室內安全行走的自動導引，採用一頂安全帽連接四台Xtion PRO LIVE裝置來獲取周遭環境的資訊，其中有一台Xtion朝下方，另外三台分別朝向前方、左方以及右方。 本論文首先提出了一個易於使用的學習方法，供系統做抓取各種環境特徵作環境地圖使用，其中包括：導引路徑、路線節點的三向深度特徵，以及使用者相對於周遭環境之位置。接著，本論文提出一個自動導引盲人的方法使盲人能夠避開障礙物以及監控周遭環境特徵。 本系統會根據三向深度特徵來判斷路徑節點來導引盲人行走當初所學習的路線地圖，在這部分，本系統不斷的利用Xtion來抓取周遭環境的深度特徵來形成三向深度特徵，用來和當初學習的資訊做比較，並利用語音的方式來導引盲人使之能夠在室內安全行走，其中在偵測盲人的位置和行走方向的方法都是利用三向深度特徵以及前方的深度影像圖來做判斷。 此外，本論文也提出偵測障礙物以及避開的方法，其中利用了moment-preserving thresholding 的技術在彩色影像圖中來偵測障礙物，並配合朝下的深度資訊來使結果更為精確，接著利用當下的三向深度特徵搭配已知的障礙物位置來導引盲人避開障礙物。 最後我們利用一實際的室內環境來測驗本論文所提出的系統，結果盲人能順利的自動修正路徑以及避開障礙物，顯示出本論文所提方法的完整性以及可行性。

A head-mounted device (HMD) system with multiple image and depth sensors for automatic guidance of a blind person to walk safely in indoor environments is proposed. The system is composed of a helmet and four Xtion PRO LIVE devices which have the capabilities of capturing image and depth data of the surrounding environment. The four Xtion devices are affixed to the helmet with one looking down and the other three facing forward, left, and right, respectively. To construct the system, at first an easy-to-use learning technique is proposed for use in the system learning stage, which has the capability of extracting various environment features to compose a path map, including a navigation path, the 3-depth feature vectors of the path nodes, and the location of the user with respect to the monitored environment. Next, an automatic guiding method for the blind person with obstacle avoidance and environment-feature monitoring capabilities is proposed. The blind person walks along the navigation path according to the node data of the path map while the proposed system monitors the path nodes by using the 3-depth feature vectors of the nodes. In this process, in each navigation cycle the system acquires depth images with the Xtion devices, extracts the environment features from the images in the form of 3-depth feature vectors, and matches them with the learned data to determine an audio instruction to guide the blind person to walk further safely. Specifically, techniques for detecting the blind-person’s position and walking direction for path correction utilizing the 3-depth feature vectors and the front depth image have been proposed. In addition, a method for obstacle detection and avoidance is also proposed, which uses several techniques, including the moment-preserving thresholding technique to find the distribution of the obstacle colors; a scheme for depth comparison between the downward depth value and the learned floor depth to detect any possible obstacle in front; and a technique of utilizing the distribution of the obstacle and the 3-depth feature vector to guide the blind person to avoid the obstacle. Good experimental results show the flexibility and feasibility of the proposed system and methods for the application of automatic guidance of the blind walking in indoor environments.