矿井智能机器人技术是煤矿智能化建设的重要支撑，而同步定位与建图技术（Simultaneous Localization and Mapping，SLAM）是实现机器人自主导航的关键。受煤矿井下低照度、弱纹理、强颠簸等复杂工况以及机器人多种运动状态的影响，传统SLAM方法容易出现定位精度下降和建图效果不佳的问题。本文基于视觉惯性传感器融合框架，针对矿井复杂场景，开展矿井机器人SLAM方法研究，主要研究内容如下：

（1）针对煤矿井下图像质量退化和特征点分布稀疏，导致SLAM系统前端有效特征不足和误匹配率高的问题，提出了一种基于矿井结构感知增强的视觉惯性里程计方法。该方法通过自适应伽马校正与灰度等间距密度均衡算法对低照度图像进行两阶段灰度重构，改善了图像的灰度分布；在此基础上，引入线特征约束，并对EDLines算法的锚点延伸策略和边缘像素链生成策略进行优化，提高了算法效率，增强了线特征对巷道结构的表达能力。实验结果表明，该方法有效提高了图像亮度及对比度，提升了点线特征数量与匹配精度，在矿井环境中表现出较强的鲁棒性与实时性。

（2）针对煤矿井下大曲率运动导致SLAM系统关键帧缺失与定位精度下降的问题，采用了一种基于视觉惯性双重约束的自适应关键帧选取策略。该策略利用惯性测量单元（Inertial Measurement Unit，IMU）感知运动过程中的视角与位置变化，进而识别机器人的大曲率运动，通过构建图像帧局部逆向索引窗口计算出自适应阈值，并结合IMU角速度阈值形成视觉与惯性双重约束条件，补全急转弯区域遗漏的关键帧。实验结果表明，该策略使大曲率运动下的关键帧提取数量增加22.40%，系统定位精度提高11.38%。

（3）针对煤矿井下复杂场景中点线特征分布不均及特征利用率低，导致位姿估计漂移的问题，构建了一种融合点线权重系数的后端非线性优化模型。该模型首先根据点线特征的提取数量与空间分布密度动态调整特征权重系数，为点线重投影误差分配自适应权重；随后将其与IMU残差结合，建立协同优化目标函数；最终基于最小二乘原理实现最优位姿解算。实验结果表明，自适应点线权重系数的引入有效提升了特征信息利用率，使系统在黑暗、弱纹理及特征分布不均场景下轨迹均方根误差降低8.2%。

（4）搭建了矿井机器人数据采集平台，开展了涵盖光照不足、粉尘干扰、特征稀疏等典型工况的模拟井下环境实验，设计了多模态运动状态测试方案，与ORB-SLAM3、EPLF-VINS等先进开源方案进行了对比分析。实验结果表明，本文系统在矿井黑暗、弱纹理环境以及机器人急转弯运动等条件下，具备更高的定位精度与鲁棒性，适用于煤矿井下复杂场景中的实时定位与建图任务。

Intelligent robot technology in mines is an important support for the intelligent construction of coal mines, and Simultaneous Localization and Mapping (SLAM) is the key to realize autonomous robot navigation. Influenced by the complex working conditions such as low illumination, weak texture, strong bumps, etc. in coal mines and the multiple motion states of robots, the traditional SLAM method is prone to the problems of decreased localization accuracy and poor mapping effect. In this paper, based on the visual inertial sensor fusion framework, we carry out the research on SLAM method for mine robots for underground complex scenes, and the main research contents include:

(1) Aiming at the problems of degraded image quality and sparse distribution of feature points in coal mine underground, which lead to insufficient effective features and high mis-matching rate in the front-end of the SLAM system, a visual inertial odometry method based on the perception enhancement of mine structure is proposed. The method utilizes adaptive gamma correction and grayscale isometric density equalization algorithm to perform two-stage grayscale reconstruction of the original image, which effectively improves the image quality; on this basis, line feature constraints are introduced into the system, and the anchor point extension strategy and the edge pixel chain generation strategy of EDLines algorithm are improved, which improves the algorithm efficiency and enhances the expressive ability of the line features for the roadway structure. The experimental results show that the odometry method effectively improves the image brightness and contrast, enhances the number of point and line features and matching accuracy, and has strong robustness and real-time performance in the mine environment.

(2) Aiming at the problem of missing key frames and degradation of localization accuracy of SLAM system due to large curvature motion in coal mine underground, an adaptive key frame selection strategy based on double constraints of visual inertia is adopted. This strategy utilizes the visual sensor and the Inertial Measurement Unit (IMU) to sense the change of view angle and position during the motion process, and then recognizes the large curvature motion of the robot, calculates the self-adaptive threshold by constructing the local inverse index window of the image frames, and combines with the IMU angular velocity threshold to form the double constraints of vision and inertia, and makes up the missing key frames of the sharp turning area. The missing keyframes are complemented to improve the quantity and quality of keyframes under large curvature motion. The experimental results show that the average number of key frames extracted by the SLAM system using this strategy in large curvature motion increases by 22.40%, and the average localization accuracy is improved by 11.38%.

(3) Aiming at the problem of drift in the position estimation caused by uneven distribution of point-line features and low feature utilization in the complex field of coal mine underground, a back-end nonlinear optimization model with the introduction of weighting coefficients is constructed. The model firstly allocates weights adaptively according to the extraction number and spatial distribution density of point and line features, then establishes a co-optimization objective function for the weighted point and line reprojection error and IMU residuals, and finally performs the optimal position solution based on the principle of least squares. The experimental results show that the introduction of adaptive point and line weighting coefficients improves the utilization rate of point and line feature information, and effectively improves the operational robustness and overall positioning accuracy of the system in dark, weak texture and uneven distribution of point and line features scenes.

(4) we build a data acquisition and processing platform for mine robots, carry out simulated underground environment experiments covering typical working conditions such as insufficient light, dust interference, and feature sparsity, and design a multimodal motion testing scheme, and compare and analyze it with advanced open-source schemes such as ORB-SLAM3, EPLF-VINS, and so on. The experimental results show that the system in this paper has higher localization accuracy and robustness under the conditions of darkness, weak texture environment in the mine, and sharp turning movement of the robot, which is suitable for real-time localization and map building tasks in complex scenarios in underground coal mines.