Sign Finding and Reading SFAR on GPU Accelerated Mobile Devices

Period of Performance: 09/30/2014 - 09/29/2015


Phase 1 SBIR

Recipient Firm

Lynntech, Inc.
Principal Investigator


DESCRIPTION (provided by applicant): The inability to access information on printed signs directly impacts the mobility independence of the over 1.2 million blind persons in the U.S. Many previously proposed technological solutions to this problem either required physical modifications to the environment (talking signs or the placement of coded markers) or required the user to carry around specialized computational equipment, which can be stigmatizing. A recently pursued strategy is to utilize the computational capabilities of smart phones and techniques from computer vision to allow blind persons to read signs at a distance using commercially available, non-stigmatizing, smart- phones. However, despite the fact that sophisticated algorithms exist to recognize and extract sign text from cluttered video input (as evidenced, for example, by mapping services such as Google Maps automatically locating and blurring out only license plate text in street-view maps) current mobile solutions for reading sign text at a distance perform relatively poorly. This poor performance is largely because until recently, smart-phone processors have simply not been able to execute state-of-the-art computer vision text extraction and recognition algorithms at real-time rates, which forced previous mobile sign readers to utilize older, simplistic, less effective algorithms. Next-generation smart-phones run on fundamentally different, hybrid processor architectures (such as the Tegra 4, Snapdragon 800, both released in 2013) with dedicated embedded graphical processing units (GPUs) and multi-core CPUs, which make them ideal for high-performance, vision-heavy computation. In this study, we propose to develop a smart-phone-based system for finding and reading signs at a distance which significantly outperforms previous such readers by implementing state-of-the-art text extraction algorithms on modern smart-phone hybrid GPU/CPU processor architectures. In Phase I, the proposed system will be developed and tested with blind users. In Phase II, feedback from user testing will be integrated into system design and the performance will be improved to permit operation in extremely challenging (such as low light) environments.