室内环境下融合点线特征的双目VI-SLAM方法

Stereo VI-SLAM method with fused point and line features in indoor environments

关键词:  同步定位与建图（SLAM）  视觉惯性  点线特征  双目相机  多尺度  非线性优化

Key words:  simultaneous localization and mapping(SLAM)  visual-inertial  point and line features  stereo camera  multi-scale  nonlinear optimization

摘要:

为解决室内环境下弱纹理场景中关键点特征稀疏、结构化场景中结构化特征应用不充分以及相机快速移动时关键帧容易跟踪失败的问题,提出了一种基于点线特征融合的双目视觉惯性SLAM方法。首先,基于EDlines线段提取方法,结合高斯图像金字塔实现多尺度线段的提取,以增强线段匹配的尺度不变性。同时,对不同尺度下的线段端点的不确定性进行建模,并结合平铺技术对线段的二进制描述符进行分块处理,从而加速线段匹配并提高线特征匹配的鲁棒性与效率。其次,优化惯性传感器的预积分模型,融合双目视觉的点特征重投影误差、线特征重投影误差以及惯性传感器的预积分约束,采用滑动窗口的非线性优化方法进行联合优化,以提高系统位姿估计精度。最后,文中在包含弱纹理、结构化以及相机快速移动等复杂环境的EuRoC数据集上进行实验。结果表明,VI-SLAM方法在EuRoC数据集上的相机轨迹均方根误差为0.031 m,平均误差为0.027 m,拥有更强的鲁棒性和更高的定位精度,尤其在弱纹理和相机快速运动场景中,定位精度优势显著。

Abstract:

To address the issues of sparse keypoint features in weakly textured indoor environments, insufficient utilization of structured features in structured scenes, and keyframe tracking failures during rapid camera movements, a stereo visual-inertial SLAM method based on the fusion of point and line features is proposed. First, the EDlines line segment extraction method, combined with a Gaussian image pyramid, is employed to achieve multi-scale line segment extraction, enhancing the scale invariance of line segment matching. Simultaneously, the uncertainty of line segment endpoints at different scales is modeled, and binary descriptors of line segments are partitioned using tiling technology to accelerate line segment matching, thereby improving the robustness and efficiency of line feature matching. Second, the pre-integration model of the inertial sensor is optimized, and a sliding window nonlinear optimization is performed by fusing the point feature reprojection error from stereo vision, the line feature reprojection error, and the pre-integration constraints of the inertial sensor, thereby improving the system’s pose estimation accuracy. Finally, extensive experiments are conducted on the EuRoC dataset which includes complex environments such as low-texture, structured scenes, and rapid camera movements. The experimental results demonstrate that the proposed method achieves a root mean square error of 0.031 m and an average error of 0.027 m on the EuRoC dataset, exhibiting stronger robustness and higher localization accuracy, especially in low-texture and rapid camera movement scenarios where the accuracy advantage is particularly significant.