A Deep-Learning-Based Floor Detection System for the Visually Impaired

The American Foundation for the Blind (AFB) has recently reported that over 25 million people in the U.S. suffer from total or partial vision loss. As a result of their visual impairment, they are constantly affected by the risk of a fall and its consequences. Most of this affected population relies on assistive technologies that allow their integration to society. Fall prevention is an area of research that focuses on the improvement of people's lives through the use of pervasive computing. This work introduces a fall prevention system for the blind and its different modules and focuses on the first module: a deep-learning approach for floor detection. A combination of convolutional neural layers and fully connected layers is used to create a network topology capable of identifying floor areas in pictures of multiple indoor environments. This task is remarkably difficult due to the complexity of identifying the patterns of a floor area in different scenarios, the noise added by the movement of the camera while walking, and the real-time nature of the system. This paper provides a general description of the fall prevention system, and a detailed description of the floor detection system. Finally, the evaluation of the proposed floor detection approach is presented: an accuracy of 92.7%, a precision of 90.2%, and a recall of 90.5% were obtained.