基于双步边缘化与关键帧筛选的改进视觉惯性SLAM方法

An approach to improve VI-SLAM based on two-step marginalization and keyframe selection method

针对目前主流视觉/惯导组合SLAM算法后端优化中计算代价大的问题,提出了一种改进边缘化策略和关键帧筛选策略的SLAM方法。首先,在系统后端优化过程中对因子图中的因子分类,通过对误差优化方程中因子类别对应的分块矩阵进行分步边缘化,将高维度矩阵分步拆解进行优化求解,提高了系统的计算效率;其次,通过增加非关键帧滑动窗口之间的图像帧约束关系,改进关键帧筛选策略,避免因视差过大导致特征点跟踪失败,从而提高了系统的稳定性与定位精度。实验表明,所提出的方法相较现有主流VI-SLAM算法在Eu Ro C数据集的平均运行速度平均提高了14.91%,且定位精度也有一定改善。

Aiming at the problem that the mainstream visual-inertial SLAM(VI-SLAM)has a high computational cost in the back-end optimization,a VI-SLAM system based on two-step marginalization and keyframe selection method is proposed.Firstly,in the process of back-end optimization,the factors in the factor graph are classified,and the block matrix corresponding to the factor category in the error optimization equation is marginalized.By decomposing the high-dimensional matrix step by step,the computational efficiency of the system is improved.Secondly,the keyframe selection strategy is improved by increasing the constraint relationship between non-keyframe sliding windows.The proposed algorithm can avoid the feature point tracking failure caused by the large parallax,and improve the stability and average accuracy of the system.Experimental results show that the computational efficiency of our method in the EuRoC dataset is improved by 14.91%compared with the state-of-art SLAM,and the positional accuracy is also improved.