This paper introduces an innovative, vision-based embedded system designed to assist visually impaired individuals by describing their surroundings using computer vision. Over 2.2 billion people around the world experience some form of vision impairment, making accessible and affordable assistive technologies crucial. Traditional aids like guide dogs and canes have limitations in providing detailed information about the environment. The proposed solution enhances the widely used YOLO (You Only Look Once) model for object detection, improving it by modifying the loss function. Specifically, it integrates CIoU (Complete Intersection over Union), focal loss, and uncertainty weights into the loss calculation, making the model more accurate and precise. These improvements lead to better results, offering more reliable assistance to visually impaired users.