Smart Object Detector System for Visually Impaired

Visually impaired individuals face many challenges in the modern world, from identifying common objects and reading, to social problems such as face recognition or picking up social cues. Most of these problems require assistance from other people or tools, such as white canes or guide dogs. However, despite being useful, these tools have limitations that can be addressed with modern technology, especially with the recent advancements in artificial intelligence and deep learning models. This paper proposes a Smart Object Detector System that aims to increase the autonomy and safety of visually impaired individuals in outdoor environments using accurate deep learning models. The system employs the YOLO architecture, exploring a set of optimizations to maximize the model's efficiency. The system was implemented on the Khadas VIM3 single computer board for power-efficient hardware acceleration, taking advantage of its neural processing unit. The optimized model running on the Khadas platform runs at 15 to 20 FPS, enough to guarantee real-time performance for the smart object detector system.