轻量卷积模型从航拍图中提取视觉地标，助力无GPS无人机导航
Compact Convolutional Segmentation for Visual Landmark Extraction in GNSS-Denied UAV Navigation
- 用空洞卷积与残差特征传递，实现高效视觉地标分割
- 在建筑分割数据集上验证可行，但需更多无人机数据优化
- 适合做无人机自主导航的前端视觉模块，尤其在信号中断时
当全球导航卫星系统（GNSS）信号受到干扰、遮挡或被恶意压制时，无人飞行器（UAV）的可靠定位变得极具挑战。在此类无GNSS环境下，机载摄像头获取的视觉信息可通过识别空间稳定且具区分性的地标来提供导航辅助。本文提出一种紧凑型卷积分割框架，用于从航拍图像中提取候选视觉地标。该模型结合全卷积处理、基于空洞卷积的空间上下文提取以及残差特征传递机制。由于缺乏专用的无人机地标数据集，本研究采用航拍建筑分割数据集作为初始评估环境。实验结果表明，所提架构可作为候选地标提取的可行前端，但未来仍需通过扩展训练、引入无人机专属数据集，并与定位或匹配算法集成以进一步提升性能。
原文摘要 · Abstract (English)
Reliable localization of unmanned aerial vehicles (UAVs) becomes challenging when Global Navigation Satellite System (GNSS) signals are degraded, blocked, or intentionally jammed. In such GNSS-denied conditions, visual information obtained from onboard cameras can provide complementary cues for navigation by identifying spatially stable and distinc tive landmarks. This study proposes a compact convolutional segmentation framework for extracting candidate visual land marks from aerial imagery. The proposed model combines fully convolutional processing with dilation-based spatial con text extraction and residual feature transfer. Since a dedicated UAV landmark dataset is not available in this study, an aerial building segmentation dataset is adapted as an initial evaluation environment. Experimental results indicate that the proposed architecture provides a feasible front-end for candidate landmark extraction, while further improvements are required through extended training, UAV-specific datasets, and integration with localization or matching algorithms.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。