摘要
针对室内弱纹理场景下特征点数量不足导致即时定位与建图(SLAM)系统跟踪丢失和重建精度差的问题,提出了一种顾及约束退化的多特征融合RGB-D SLAM算法。为了充分利用线和平面特征对位姿估计的约束,分别建立了线和平面误差方程,并通过对海森矩阵进行特征值分解,定量分析了线和平面特征位姿约束的退化情况,建立了顾及约束退化的多特征融合目标优化函数。此外,基于曼哈顿世界假设,建立了曼哈顿坐标系,充分利用曼哈顿世界假设的优势,对旋转矩阵的“零漂移”进行估计,以提供准确的初始值支持平面匹配和位姿优化。实验结果表明,引入线和面特征建立光束法方程后,所提出的方法在弱纹理数据集ICL-NUIM上的轨迹精度相较于基准的ORB-SLAM2平均提升了37.5%,有效改善了SLAM系统在弱纹理场景中的轨迹精度。
A multi-feature fusion RGB-D SLAM algorithm considering constraint degradation is proposed to solve the problem of tracking loss and reduced reconstruction accuracy in SLAM system due to insufficient number of feature points in indoor scenes with weak texture.In order to exploit the constraints provided by line and plane features for pose estimation,error equations for lines and planes are established respectively.By performing an eigenvalue decomposition of the Hessian matrix,the degradation of pose constraints imposed by line and plane features is quantitatively analyzed,paving the way for the establishment of a multi-feature fusion objective optimization function that considers constraint degradation.In addition,by exploiting the Manhattan World assumption,a Manhattan coordinate system is established to estimate the zero drift of the rotation matrix,providing accurate initial values to support plane matching and pose optimization.Experimental results show that after introducing line and plane features to establish the bundle adjustment equation,the proposed method improves the trajectory accuracy on the low-texture dataset ICL-NUIM by 37.5%compared to the benchmark ORB-SLAM2,effectively improving the trajectory accuracy of SLAM systems in weakly textured environments.
作者
王西旗
毕京学
杨尚帅
WANG Xiqi;BI Jingxue;YANG Shangshuai(School of Transportation Engineering,Shandong Jianzhu University,Jinan 250101,China;School of Surveying Mapping and Geographic Information,Shandong Jianzhu University,Jinan 250101,China;The First Topographic Survey Team of the Ministry of Natural Resources,Xi'an 710054,China)
出处
《导航定位与授时》
CSCD
2024年第5期53-65,F0002,共14页
Navigation Positioning and Timing
基金
国家自然科学基金(42001397)。
关键词
即时定位与建图
多特征融合
室内弱纹理场景
曼哈顿世界假设
RGB-D相机
约束退化
Simultaneous localization and mapping(SLAM)
Multi-feature fusion
Indoor weak texture scenes
Manhattan World assumption
RGB-D camera
Constrained degradation
