Visual-inertial SLAM is essential in various fields, such as AR/VR, uncrewed aerial vehicles, industrial robots, and autonomous driving. The fusion of a camera and inertial measurement unit (IMU) can make up for the shortcomings of a signal sensor, which significantly improves the accuracy and robustness of localization in challenging environments. Robust tracking and accurate inertial parameter estimation are the basis for the stable operation of the system. This article presents PLE-SLAM, an entirely precise and real-time visual-inertial SLAM algorithm based on point-line features and efficient IMU initialization. First, we introduce line features in a point-based visual-inertial SLAM system. We use parallel computing methods to extract features and compute descriptors to ensure real-time performance. Second, the proposed system estimates gyroscope bias with rotation pre-integration and point and line observations. Accelerometer bias and gravity direction are solved by an analytical method. After initialization, all inertial parameters are refined through maximum a posteriori (MAP) estimation. Moreover, we open a dynamic feature elimination thread to improve the adaptability to dynamic environments and use CNN, bag-of-words and GNN to detect loops and match features. Excellent wide baseline matching capability of DNN-based matching method and illumination robustness significantly improve loop detection recall and loop inter-frame pose estimation. The front-end and back-end are designed for hardware acceleration. The experiments are performed on public datasets, and the results show that the proposed system is one of the state-of-the-art methods in complex scenarios.

翻译：视觉惯性SLAM在AR/VR、无人机、工业机器人和自动驾驶等多个领域至关重要。相机与惯性测量单元的融合可以弥补单一传感器的不足，显著提升复杂环境下定位的精度与鲁棒性。鲁棒的跟踪和准确的惯性参数估计是系统稳定运行的基础。本文提出PLE-SLAM，一种完全精确且实时的基于点线特征与高效IMU初始化的视觉惯性SLAM算法。首先，我们在基于点的视觉惯性SLAM系统中引入线特征。采用并行计算方法提取特征并计算描述子以确保实时性。其次，本系统通过旋转预积分以及点和线的观测值估计陀螺仪偏置。加速度计偏置和重力方向通过解析方法求解。初始化完成后，所有惯性参数通过最大后验估计进行优化。此外，我们开辟动态特征剔除线程以提高对动态环境的适应能力，并利用CNN、词袋和GNN进行闭环检测与特征匹配。基于深度神经网络的匹配方法具有出色的宽基线匹配能力和光照鲁棒性，显著提升了闭环检测召回率和闭环帧间位姿估计精度。前端与后端均针对硬件加速设计。在公开数据集上的实验结果表明，本系统在复杂场景下达到了最先进水平之一。