Discipline: Computer Sciences and Information Management
Subcategory: Computer Science & Information Systems
Osniel Quintana - University of South Florida
There are over a 100 million people that suffer from injuries due to falls, with the blind and elderly being the most affected. Researchers from the medical field have tried to create a solution for these problems, but so far, they haven’t been able to engineer a product able to eradicate this issue. Currently, there is no affordable solution available in Computer Science since in order to detect a fall, the program needs to be able to filter accurate real-time data while maintaining an energy-efficient fall detection unit. This causes an inverse proportionality between scalability and affordability. The aim of this project is to aid a visually impaired person, or an elder, during his daily commute by preventing him from falling. The solution proposed uses Machine Learning and Computer Vision to accurately determine how far an object is from the user’s path. Using Artificial Neural Networks (ANN), specifically deep neural networks, the solution will be scalable yet affordable. The system will be based on an Android application that uses the neural network to analyze color images and determine how far an object is from the user. Currently the system is able to determine the distance from the user to objects in the environment with a 92 Symmetric Mean Percentage Error. This high percentage error was due in great part to a faulty infrared sensor used to collect the images that would become the ground truth. This research will be reproduced using an Intel RealSense camera to collect the images again.
Not SubmittedFunder Acknowledgement(s): This research has been partially supported by the National Science Foundation under grants No. 1458928 and No. 1645025, An REU Site on Ubiquitous Sensing
Faculty Advisor: Miguel Labrador, labrador@csee.usf.edu
Role: I was the only person conducting this research. I developed the algorithm to capture both the infrared images and the color images, created the neural network model, trained and evaluated the model and finally I determined the main cause of error during the training phase.