针对目前采场支架姿态感知中惯导方法存在漂移误差大、解算精度低，以及视觉方法存在相机易受粉尘与设备遮挡而位姿识别误差大等问题，提出了一种基于视觉−惯导信息融合的采场支架姿态感知方法。首先将四特征点红外标靶固定于支架底座凸台，将双目相机分别固定于支架顶梁与掩护梁，采用基于Canny−最小二乘法的靶标识别方法和基于四特征点的BA−PnP算法解算顶梁、掩护梁相对底座的俯仰角、横滚角。然后将惯性测量单元固定于液压支架顶梁、掩护梁、底座，通过惯性测量单元中MEMS陀螺仪和加速度计的互补滤波方法解算顶梁、掩护梁、底座在世界坐标系下的俯仰角、横滚角。最后将视觉系统解算的姿态角与惯导解算的姿态角进行扩展卡尔曼滤波多源信息融合，利用视觉信息的低频稳定性抑制惯性测量单元的累计误差，得到精确的采场支架姿态。采用基于视觉、惯导和视觉−惯导信息融合3种支架姿态感知方法进行对比实验，结果表明：① 初始静止状态下，3种方法的精度均较高，但随着支架运行循环次数增加，基于视觉、惯导的感知结果逐渐偏离真值。② 基于视觉、惯导和视觉−惯导信息融合方法的顶梁相对底座的俯仰角感知均方根误差分别为0.201，0.190，0.081°，掩护梁相对底座的俯仰角感知均方根误差分别为0.340，0.297，0.162°。③ 基于视觉−惯导信息融合方法解算的液压支架立柱伸缩长度的均方根误差为13.682 mm，满足现场需求。基于视觉−惯导信息融合的采场支架姿态感知方法可为液压支架智能化控制提供更准确的姿态参数。

To address the issues of large drift errors and low calculation accuracy in inertial navigation methods, as well as significant posture recognition errors in visual methods due to camera interference from dust and equipment obstructions, a method of support posture perception in mining face based on visual-inertial information fusion was proposed. First, four feature points of infrared targets were fixed to the base platform of the support, and binocular cameras were fixed to the support top beam and shield beam. A target recognition method based on Canny and least squares, along with a BA-PnP algorithm based on four feature points, was used to solve the pitch and roll angles of the top beam and shield beam relative to the base. Then, an inertial measurement unit (IMU) was fixed to the hydraulic support top beam, shield beam, and base. The complementary filtering method of the MEMS gyroscope and accelerometer in the IMU was used to solve the pitch and roll angles of the top beam, shield beam, and base in the world coordinate system. Finally, the posture angles calculated by the visual system and the inertial navigation system were fused using the extended Kalman filter for multi-source information fusion. The low-frequency stability of the vision information was used to suppress the accumulated errors of the IMU, resulting in accurate posture perception of the mining support. Three methods for support posture perception, based on vision, inertial navigation, and visual-inertial information fusion, were compared in experimental results. The findings showed that: ① In the initial stationary state, all three methods had high accuracy, but as the support operation cycles increased, the vision-based and inertial navigation-based results gradually deviated from the true values. ② The root mean square errors (RMSE) of pitch angle perception for the top beam relative to the base were 0.201°, 0.190°, and 0.081° for the vision-based, inertial navigation-based, and visual-inertial information fusion methods, respectively. For the shield beam relative to the base, the RMSE of pitch angle perception were 0.340°, 0.297°, and 0.162°, respectively. ③ The RMSE of the hydraulic support column extension length calculated by the visual-inertial information fusion method was 13.682 mm, meeting on-site requirements. The visual-inertial information fusion-based support posture perception method could provide more accurate posture parameters for the intelligent control of hydraulic supports.