摘要
针对灾难环境可能造成全局定位系统的失效及可见光图像的退化,以及传统的基于计算机视觉的重定位算法因图像特征点不足而成功率较不高的问题,提出了一种基于语义地图的无人机重定位方法。该方法依赖RGB-D图像来识别并构建受灾环境中的路标点,通过与先验地图间的路标点匹配,进而优化求解得到无人机的相对位姿。通过减少对目标识别网络潜在的广义物体识别能力的抑制,得到图像中的高级别特征点,有效解决了因特征点不足而难以重定位的问题。基于广义物体重建的路标点,提出了一种快速的路标点检索与匹配方法。实验结果表明,与基于目标识别网络的路标点构建方法相比,本方法能在未知环境中重建更丰富的路标点,并能有效地基于这些路标点进行重定位。在图像退化的灾难场景中,本方法展现出比目前被广泛使用的图像检索方法更高的召回率和鲁棒性。
In response to the potential failure of global positioning systems in disaster environments and the degradation of visible light images,as well as the low success rate of traditional computer vision-based relocalization algorithms due to insufficient image feature points,a semantic map-based drone relocalization method for unmanned aerial vehicle(UAV)is proposed.This method relies on RGB-D images to identify and construct landmark points in the disaster-affected environment.These landmark points are then matched with prior maps to optimize and estimate the relative pose of the drone.By reducing the suppression of the potential general object recognition capability within object recognition networks,high-level feature points in the image are obtained,effectively addressing the problem of difficult relocalization due to insufficient feature points.Building on the generalized objectbased reconstruction of landmark points,an efficient method for retrieving and matching these points is proposed.Experimental results demonstrate that the approach can reconstruct a richer set of landmarks in unknown environments and effectively utilize them for localization,compared to other object recognition-based landmark point construction methods.In disaster scenarios with image degradation,this method exhibits higher recall rates and robustness than widely used image retrieval methods.
作者
黎容熙
唐家成
胡天江
LI Rongxi;TANG Jiacheng;HU Tianjiang(School of Aeronautics and Astronautics,Sun Yat-sen University,Shenzhen 518107,China;School of Artificial Intelligence,Sun Yat-sen University,Zhuhai 519080,China)
出处
《中山大学学报(自然科学版)(中英文)》
北大核心
2025年第3期109-118,共10页
Acta Scientiarum Naturalium Universitatis Sunyatseni
基金
广东省重点领域研发计划(2024B1111060004)。
关键词
城市火灾
无人机
重定位
语义地图
回环检测
urban fire
UAV
relocalization
semantic map
loop closure detection
