Abstract
Computer vision tasks such as image segmentation, object detection, and face recognition have been crucial in developing assistance systems for visually impaired individuals. Among these, image segmentation plays a vital role in helping them navigate safely. However, this task is more complex as it requires detailed spatial information. In this article, we propose a novel panoptic segmentation framework that serves as the foundation for an effective pathfinding system, combining robust collision avoidance with high performance. Our contribution includes a single-stage instance segmentation method built on a ResNet101-FPN encoder-decoder architecture. Additionally, we created a customized panoptic labeled dataset to meet the specific needs of visually impaired individuals, aiming to support future integration with real-time feedback in visual prostheses. We evaluate our model both qualitatively and quantitatively using the Panoptic Quality (PQ) metric. Results show that our method surpasses recent panoptic segmentation techniques, achieving a PQ score 4.092 points higher. It also outperforms existing pathfinding systems, demonstrating greater accuracy and efficiency under varied weather conditions.