In this paper, we implement a deep learning model to detect obstacles on roads for the blind individuals and autonomous delivery robots. We applied various methods to get model with high accuracy. Firstly, we used YOLOv7-tiny as a deep learning model, which demonstrated excellent performance due to its efficient architecture. Secondly, we used median frequency balancing to solve class imbalance in the dataset, resulting in an increase of 0.4 in mAP. We also conducted experiments with different bounding box regression losses, such as GIoU, DIoU, and CIoU, as well as classification losses, such as FL, QFL, and VFL, to improve the efficient training and performance of the model. As a result, CIoU, which considers all important factors in bounding box regression, and VFL, which effectively addresses foreground-background class imbalance, showed the best performance with 68.7 mAP, surpassing the baseline by over 3% in mAP.
