MY VISION A MOBILE APPLICATION TO INCREASE SAFETY OF VISUALLY IMPAIRED PEOPLE

One of the consequences of having impaired vision is being uncomfortable about safety
while moving around or travelling independently. Individuals with visual impairment
have many difficulties in self-navigation in unfamiliar outdoor environments. The aim of
this project is to develop a mobile application as an assistive technology that consists of
obstacle detection and obstacle warning for the visually impaired to safely navigate in
outdoor environment.
The following objectives have been identified to fulfil the aim of this project. I)
To identify suitable machine learning models for object detection in real-time using
mobile application. II) To identify suitable methods for obstacle warning in real-time
using text-to-speech functionality. III) To evaluate the performance of the developed
mobile application in terms of detection and warning accuracy and user satisfaction level.
In order to fulfill the objective of the project, Literature review has been conducted
to find the most suitable object detection method for the mobile application and study the
kind of commands for obstacles warning. Besides, the mobile application will be using
the text-to-speech technologies for the obstacles warning. Furthermore, test will be
conducted during the model training process to evaluate the mean average precision and
loss of the model.
Based on the Literature review, SSD MobileNet is the most suitable method for
mobile application object detection. Although Faster R-CNN has the higher accuracy,
however the method is required high computation usage, therefore it is not suitable for
mobile application. The result and discussion have presented the distance of detection,
detection output which included audio and visual display, user testing and loss and mean
average precision.
The mobile application is able to find the right boxes for detected object and giving
out appropriate obstacles warning to the user. However, due to overfitting of the model,
the detection might experience inaccurate if detect newly unseen data. Lastly, the author
has proposed several method to increase the performance of the model and effectiveness
of the mobile application.