Object Recognition for Visually Impaired Pedestrian Navigation Using Lightweight Yolov8n Model

Abstract

Real-time object detection and recognition based on computer vision and AI play an important role in assisting individuals with visual impairment. However, existing related tools have limitations in immediate auditory feedback to the users. Thus, this paper employed YOLOv8n model to address the limitations. The model is selected for its lightweight design, fast inference speed, and reliable accuracy, along with text-to-speech. A mobile application was developed to integrate detection cameras that features real-time audio feedback to describe detected objects in pedestrian surroundings. The compact YOLOv8n model maintained its performance across varied lighting conditions while operating efficiently on limited resources devices. Testing and evaluation demonstrated that the system was able to detect real-time common objects and deliver prompt audio feedback with overall accuracy of more than 70%. The findings highlighted the potential of combining machine learning models with cross-platform mobile framework to create assistive tools that enhance spatial awareness for visually impaired users. Overall, this real-time object detection and recognition achieved its objectives to support and empowers visually impaired users in navigating pedestrian environments independently by offering a functional and real-time object detection and recognition solution.