Deep-learning-based visual aid for low vision

Abstract

Low vision significantly impacts daily navigation, affecting an estimated 295 million people globally (Bourne et al., 2021). While Head-Mounted Displays (hereafter HMDs) have been used throughout the latest decades to mitigate these limitations, to our knowledge, no low-cost existing system utilizes real-time object detection within an HMD for navigation assistance specifically designed for the visually impaired. This letter presents the development and evaluation of a real-time navigation assistance device with automatic detection and highlighting of visual Points of Interest over the visual field, deployed on a resource-constrained embedded platform. We leverage state-of-the-art YOLO models (v8, v9, and v10), optimized for execution on a Raspberry Pi 5. Using optimized embedded systems inference engines, we investigate the tradeoffs between accuracy, speed, and power consumption. Using a custom-built dataset and publicly available benchmark files, our experimental results demonstrate the feasibility of achieving real-time performance with acceptable accuracy on a low-cost, portable device, enabling practical semantics-aware assistive technology for the visually impaired.