Object Recognition Mobile App for Visually Impaired User

• Main Author: Won, Wei Cheng
• Format: Final Year Project / Dissertation / Thesis
• Published: 2021
• Subjects: QA76 Computer software
• Online Access: http://eprints.utar.edu.my/4098/1/1803164_FYP_report_%2D_WEI_CHENG_WON.pdf; http://eprints.utar.edu.my/4098/

Vision is one of the most important human senses and it plays a critical role in understanding the surrounding environment. However, millions of people in the world experience visual impairment. These people face difficulties in their daily navigation since they are unable to see the obstacles in their surroundings. Despite there are many options such as white canes and different advanced technologies to help visually impaired people when navigating, some of the options are unreliable, expensive, and hard to access. Hence, a mobile application is proposed to help visually impaired people to recognise objects in their surroundings using real-time object detection and object recognition techniques. This project also has applied transfer learning on multiple pretrained models to train the models that are able to classify 40 classes of objects. The performance of the trained models is compared to select a suitable model to be implemented in the mobile application. The Evolutionary Prototyping Model is the development methodology adopted in this project. It involves developing the application in a series of iterations and refining the application based on feedback collected in each iteration. A literature review was conducted on similar existing mobile applications to understand the machine learning framework used for the implementation of object detection and recognition, and also identify the important features and workflow within the application. Finally, an Android-based mobile application was developed successfully and passed all testing. In conclusion, this project has helped visually impaired people to determine the objects in their surrounding in a more cost-effective, accessible and reliable way. They are being informed of the names and directions of the detected objects in the surroundings through voice feedback without requiring any network connection or photo capturing.