حسن سلام عبدالامير
One of the most important senses for living is vision. Millions of people living in this world deal with visual impairment. These people are exposed to many problems in their daily life.
The blind and visually impaired people cannot perform daily tasks by themselves. To perform these tasks, such as finding the objects in their indoor and outdoor environment, they depend on other people. Those people maybe untrusted, or they rely on other senses such as hands, and this interaction between visually impaired people and objects maybe dangerous or even lethal.
The objective of this thesis is to change the visual world into an audio world by using the smart eyeglasses’ system for notifying the blind people about the static objects in their indoor environment.
The Smart Eyeglasses is an intelligent system similar to the human eyes in its function, helps the Blind and Visually Impaired People inside the room to detect and recognize the static objects easily and independently at less effort.
The smart Eyeglasses’ system consists of USB camera, glasses, Raspberry Pi, Power bank and earphones. The USB camera which is placed on the glasses is used for capturing an image via using Open source Computer Vision (OpenCV). Then, this image is sent to the Raspberry Pi which is a small device similar to the brain of human in its function, used to analyze the image for detecting the objects in the image by using long complex algorithm known as the deep Learning algorithm. When the objects are detected and recognized, the sound in Arabic language will be generated and sent to the earphones which are placed on the ear of visually impaired people.
The proposed smart Eyeglasses’ system can detect and recognize eight types of objects (TV, Person, Bottle, Chair, Table, Laptop, Printer and water dispenser).
The code inside the Raspberry Pi is written by python programming language (Python 2). Google Drive cloud storage is used to save the dataset for the purpose of training the deep neural network. The dataset is consisted of 1960 image (eighty-five percent of total images is used for training the neural network while the remainder is used as testing images). Tensorflow is used to build the proposed model. Google Colab cloud service is used to train the deep neural network online. Playsound library is used for the purpose of converting text into sound. YOLO deep learning algorithm is used for detecting and recognizing the objects and it achieved an accuracy of 98%.
In addition, this thesis demonstrated that YOLOv3 outperforms YOLOv3-Tiny in detection accuracy.
