Object Detection for Assistive Technology: A Comprehensive Real-World Performance Evaluation using a Dedicated Framework

This study explores the feasibility of using object detection technology to aid visually impaired individuals by providing real-time information about the surrounding environment. Recognizing the challenges faced by millions worldwide with visual impairments, this research investigates the potential of object detection systems to provide crucial, real-time environmental awareness. A custom framework was developed to facilitate data collection and analysis, using data from both real-time device operation and diverse image datasets. This approach allowed for a comprehensive evaluation of performance accuracy and reliability in realistic, real-world settings. The research assesses the capabilities and limitations of object detection for assistive technology, focusing on its potential to accurately identify and locate objects critical for navigation and interaction. Two representative models were evaluated: YOLOv11, chosen for compute-intensive tasks, and KidBright μAI, a representative platform for edge AI deployment. YOLOv11, running on Raspberry Pi 5, demonstrated strong performance with an average F1-score of 0.7762, precision of 0.9380, and recall of 0.8100, but required 3.40 J per inference. KidBright μAI, designed for edge computing, achieved an F1-score of 0.5992, precision of 0.5621, and recall of 0.6383, with a significantly lower energy consumption of 0.21 J per cycle. In user trials, the assistive device, utilizing YOLOv11 due to its higher precision and lower false positives, improved object localization success rates from 50% to 85.7<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> for visually impaired participants. The findings highlight the potential of object detection for assistive technology, with YOLOv11 offering higher accuracy and KidBright μAI providing a more energy-efficient solution.