A LaMa-based Panoramic Image Restoration Framework for Enhanced Omnidirectional Obstacle Detection in Low-Altitude UAVs

Omnidirectional perception is a critical component for enabling autonomous navigation in low-altitude UAVs. This paper presents a novel framework that integrates LaMa-based image restoration with YOLOv8 object detection to enhance obstacle detection in panoramic images captured by fisheye cameras. In UAV operations, panoramic images often suffer from stitching artifacts and lens distortions, which can impair detection accuracy. To address this challenge, we propose an artifact-aware mask generation method to identify and restore corrupted regions using the LaMa inpainting model. The restored panoramic images are then processed by YOLOv8, resulting in accurate and robust detection of obstacles such as trees, power lines, and buildings. Experimental results, based on synthetic and real-world UAV datasets, show significant improvements in detection performance, including a notable increase in mAP, precision, recall, and F1-score compared to baseline methods. Our system demonstrates real-time processing capabilities on embedded platforms, providing a reliable solution for omnidirectional obstacle detection in UAVs. The proposed framework lays the foundation for future advancements in UAV-based environmental perception and autonomous navigation.