基于双重初始化和分级优化的改进视觉惯性SLAM方法

An approach to improve visual-inertial SLAM method based double initialization and hierarchical optimization

针对传统视觉惯性SLAM初始化阶段收敛速度慢、准确性低以及抗干扰能力较弱,实时优化过程局部轨迹累计误差高、后端优化精度低等问题,提出一种双重初始化和关键帧分级优化的改进视觉惯性SLAM方法。在视觉惯性导航系统(VINS)初始化基础上构建观测样本情形评估函数,评估状态误差信息矩阵,判断初始化算法终止条件,缩短初始化时间,提高系统鲁棒性。在后端优化阶段建立紧耦合图优化模型,在滑动窗口优化之前以共视关系分级关键帧,优化强共视关键帧,消除局部累计误差,提高后端优化精度。在EuRoC数据集上进行的实验结果表明,所提方法较VINS-mono漂移误差减少约34%,时间效率提高11.19%,建图精度得到明显提升。

Aiming at the problems of slow convergence speed,low accuracy and weak anti-interference ability in the initialization phase of traditional visual inertial navigation SLAM,high cumulative error of local trajectory during real-time optimization,and low accuracy of back-end optimization,a two-step initialization and key frame improved visual inertial SLAM method for hierarchical optimization is proposed.On the basis of VINS initialization,the evaluation function of the observation sample situation is constructed,the state error information matrix is evaluated,the termination condition of the initialization algorithm is judged,the initialization time is shortened,and the system robustness is improved.In the back-end optimization stage,a tightly coupled graph optimization model is established.Prior to the sliding window optimization,the key frames are classified according to the common view relationship,and the strong common view key frames are optimized to eliminate local cumulative errors and improve the accuracy of the back-end optimization.The experimental results on the EuRoC data set show that the proposed method reduces the drift error of VINS-mono by about 34%,improves the time efficiency by 11.19%,and improves the mapping accuracy significantly.