個人型之盲人引導輔助系統
摘要
　相較於其他感官障礙者，視覺障礙者在生活自理上有更多的不方便，且相關的輔助工具進步的非常緩慢，像是白手杖已經使用超過 100 年了，地位依然屹立不搖，而導盲犬普及也是受到重重的障礙， 因此，這幾年來使用電腦視覺的方式增加視障者生活自理能力的相關研究越來越蓬勃發展。 　　本論文使用 Kinect 深度攝影機，建立視障者生活輔助的系統，系統主要有兩大功能。 (1)利用 Kinect 提供的路面法向量，配合侵蝕與擴張演算法，重建路面資訊，進而偵測路面上的資訊，如路長、路面上障礙物等，提供視障者行走時能多一分路面的資訊。 (2)輔助視障者尋找日常生活用品，首先利用卷積神經網路對日常生活中的物品建立辨識的模型，實際應用時，利用畫面切割與統計方式找出物品在彩色影像中的位置，以提供視障者能快速找到物品。 　　本系統所使用之路面切割偵測演算法與實際大約只差 1.2%，能有效切割出屬於路面的區域，並且補足 Kinect 深度資訊誤差所造成的影響。 物件偵測的三種情境各準備一萬張影像，最壞的情況也只有不到 5%的機率找不到物品，足以證明本論文所使用之方法能有效地找出目標物品。

關鍵字： Kinect、視覺障礙者、卷積神經網路、物品偵測

A Personal Guidance Aid System for the Blind
Abstract
　Compared to other disabilities, people with visual impairment are more inconvenient to take care themselves. However, the development in blindness guidance aids is slow, for example, white cane has been used for more than a hundred years, but there appears no new aids can completely replace it. Guide dogs are another option, but it is not easy to use widely. Therefore, more and more researches of aid system using computer vision have been published. 　　In this paper, we use Kinect as the depth sensor to build a guidance aid system for blinds. The system has two features. (1) Rebuild the path information basing on the normal vector provided by Kinect with the algorithm of erosion and dilation, which could retrieve the length of road and the height of obstacle (if any), helping the visually impaired to recognize the walking path. (2) Provide the assistance for the visually impaired to find daily necessities. First, our system trains identification model from convolutional neural networks, which will finally applied for the recognition in a series of segmented images, and get the item locations from the statistical results. That could help the visually impaired quickly find those items. 　　The accuracy of the floor extraction algorithm used in the system is about 98.8%, which means it could fix the errors read from Kinect in the depth information and could efficiently extraction the regions belong to the floor. For the object detection tests, we prepared ten thousand pictures for each of the three iii conditions, and we got only less than 5% of chances not to find the target object, which certainly proves that the method we used could efficiently identify the target object. Key Words: Kinect, visual impairments, convolutional neural networks, item detection