行走區域標示及危險狀況判別之盲人輔助系統

Abstract

在本論文中，我們提出一套以視覺為基礎的盲人輔助系統。本系統採用了資料庫為主的架構，首先具有正常視力之輔助者事先於盲人經常行走之區域，建立360度全景之資料庫，並對資料庫圖片事前標籤重要的物體，像是人行道及道路；然後從全景資料庫中尋找與使用者前方視野最相近的區域；接著把此最接近之區域的標籤利用兩張影像之間的對應，產生目前環境的標籤；最後利用推論的標籤來判斷盲人目前的處境是否危險。只要使用者處於有設置資料庫的環境下，本系統能幫助盲人辨識可行走的區域，達成輔助盲人安全行走之目標。

In this thesis, we propose a data-driven system that assists blind people to walk safely on the sidewalk. In our system, an assistant with normal vision is asked to create the database for the places where the blind user usually visits. At each sampling spot of these places, the assistant takes a few photos that cover different viewing directions around the sampling spot to create a panorama image. After the installation of the database, the blind user is equipped a camera while he or she is walking around these places. For each captured image by the camera, the proposed system finds the most similar panoramic part in the database to identify the location and the orientation of blind user. With an image-to-image matching to warp the labels from the matched panoramic part to the captured image, our system can roughly infer the labeling of the contents within the captured image. Finally, based on the inferred labels, our system can identify situations that could be dangerous to the blinds.