基于点线特征融合的延迟边缘化视觉惯性SLAM方法

Delayed Marginalization Visual Inertia SLAM Method Based on Point and Line Feature Fusion

摘要:

针对单一传感器SLAM技术在复杂环境中存在精度低、感知退化、可靠性差，导致无法准确估计摄像机轨迹的问题，提出一种基于点线特征融合的延迟边缘化视觉惯性SLAM算法(DM-VI-SLAM)。首先，采用因子图优化模型，提出以IMU为主系统，视觉为辅系统，通过引入辅系统观测因子约束IMU主系统偏差，并接收IMU里程计因子实现运动预测与融合的全新结构。其次，在前端加入点线特征，设计一种基于线段中点的特征匹配方法，在后端加入滑窗机制实现历史状态信息回溯，并构建非线性联合优化问题，提升匹配精度。最后，为加速求解，引入一种延迟边缘化策略，允许重新推进延迟因子图，进而产生新的和一致性的线性化点更新边缘化。通过与典型SLAM算法进行比较，并在EuRoC公共数据集上和真实场景中验证算法有效性，实验结果表明在复杂高速运动场景和低特征纹理场景中本文算法均具有更高的精度和可靠性。

Abstract:

A delayed edge based visual inertial SLAM algorithm (DM-VI-SLAM) based on point line feature fusion was proposed to address the issues of low accuracy, perceptual degradation, and poor reliability of single sensor SLAM technology in complex environments, which made it difficult to accurately estimate camera trajectories. Firstly, a factor graph optimization model was employed, proposing a novel structure that taked the inertial measurement unit (IMU) as the primary system and vision as the auxiliary system. This structure introduced auxiliary system observation factors to constrain the biases of the IMU primary system and receiving IMU odometer factors to achieve motion prediction and fusion. Secondly, by adding point and line features in the front-end, a feature matching method based on the midpoint of a line segment was designed. A sliding window mechanism was added in the back-end to achieve historical state information backtracking, and a nonlinear joint optimization problem was constructed to improve matching accuracy. Finally, to accelerate the solution, a delayed marginalization strategy was introduced that allowed for the readvancement of the delay factor graph, thereby generating new and consistent linearization points to update the marginalization. By comparing with typical SLAM algorithms and verifying their effectiveness on EuRoC public datasets and real scenes, experimental results showed that the proposed algorithm had higher accuracy and reliability in complex highspeed motion scenes and low feature texture scenes.