Object Detection for Impaired Visual Assistance Using Transfer Learning and IoT-Raspberry Pi

Abstract

Navigating outdoor environments poses substantial challenges for blind and visually impaired people, limiting their ability to move independently and safely. This thesis presents a novel AI-based system designed to enhance mobility for visually impaired users by providing real-time object detection and depth sensing. Utilizing deep learning techniques and the YOLOv8 object detection algorithm, the system is implemented on embedded systems with the Raspberry Pi 4 and integrated with a 3D camera to assess the spatial proximity of detected objects. The custom WOTR (walk on the road) dataset developed for this project, tailored to the needs of visually impaired individuals, ensures high accuracy in object detection and depth estimation. The system delivers real-time audio feedback, offering practical guidance for non-controlled outdoor assistive navigation. Comprehensive testing in various outdoor settings demonstrates the system’s effectiveness in detecting objects, estimating their depth, and providing timely feedback. The portability and cost-effectiveness of the Raspberry Pi 4 make this solution accessible to a wide audience, potentially improving the quality of life for visually impaired individuals by enabling safer and more confident navigation. This work advances the field of assistive technologies, offering a practical tool that empowers blind and visually impaired individuals to navigate outdoor spaces with greater ease and independence.