针对煤矿井下环境复杂，煤矸石大小不一、形状各异，存在被煤泥严重包裹等问题。一般的视觉识别难以实现复杂条件下煤矸石的检测，射线检测技术对煤矸石识别率高，但又存在深度信息缺失。因此，本文在团队研发的多机械臂煤矸分拣机器人平台和已获取的煤矸石X射线图像的基础上，重点研究X射线图像与双目视觉信息融合的煤矸石抓取特征提取方法。

针对煤矸分拣机器人抓取特征提取问题，构建了X射线+双目视觉的煤矸石抓取特征提取系统。系统是以煤矸石X射线图像为模板，融合双目视觉信息对目标矸石进行匹配和抓取特征提取，其可以为煤矸分拣机器人抓取矸石提供可靠的数据支撑。

针对X射线和双目视觉图像的目标矸石匹配问题，提出一种基于轮廓HU不变矩的矸石目标快速匹配方法。对获取的矸石X射线图像通过二值化、形态学以及边缘检测等处理，提取矸石的最小封闭轮廓；并以此作为匹配模板进行输入，由HU不变矩原理生成描述轮廓信息的特征向量；根据图像搜索策略和相似性函数确定的阈值在煤矸混合视觉图像中进行搜索，最终通过匹配结果成功定位到X射线识别下的目标矸石，完成了从X射线信息到双目视觉信息的传递。

针对煤矸石抓取特征提取问题，提出了用点云处理的方式生成矸石最优抓取立方体的方法。分析了可以滤除周围环境及背景等场景点云信息的预处理方法，选用直通滤波器解决了场景点云信息对目标矸石点云的干扰问题；选用体素滤波器精简了目标矸石的点云，降低了点云数据量，提高了后续点云处理的效率；选用鲁棒性更强的半径滤波器解决了点云噪声的干扰问题；比较分析了不同的点云分割方法，使用DBSCAN聚类算法完成了目标矸石点云数据的分割；使用最小方向包围盒OBB算法完成了矸石最优抓取立方体的构建，提取到了矸石抓取特征参数。

最后，针对以上提出的融合X射线图像和双目视觉信息的煤矸石抓取特征提取方法进行了实验验证，设计了整个实验方案，进行了煤矸石抓取特征提取静态实验、动态实验以及煤矸石抓取实验，结果表明，通过本文中目标矸石抓取特征提取方法为机械手在抓取矸石时提前制备好矸石抓取特征参数，可以明显地提高机械手抓取矸石的成功率和效率。

 In view of the complex underground environment of coal mine, coal gangue varies in size and shape, and is seriously wrapped by coal slime. It is difficult to detect coal gangue under complex conditions with general visual recognition. However, the recognition rate of coal gangue with ray detection technology is high, but there is lack of depth information. Therefore, based on the multi-arm coal and gangue sorting robot platform developed by the team and the obtained coal gangue X-ray images, this paper focuses on the coal gangue grabbing feature extraction method based on the fusion of X-ray images and binocular visual information.

A coal gangue capture feature extraction system based on X-ray + binocular vision was constructed to solve the problem of capture feature extraction of coal gangue sorting robot. The system is based on the coal gangue X-ray image as the template, and combines binocular visual information to match and extract the target gangue features, which can provide reliable data support for coal and gangue sorting robot to capture gangue.

Aiming at the problem of gangue target matching in X-ray and binocular vision images, a fast gangue target matching method based on contour HU invariant moment was proposed. The obtained X-ray images of gangue were processed by binarization, morphology and edge detection to extract the minimum closed contour of gangue. The HU invariant moment principle is used to generate feature vectors describing contour information. According to the image search strategy and the threshold value determined by the similarity function, the coal and gangue mixed visual images were searched, and finally the target gangue under X-ray recognition was successfully located through the matching results, and the transmission from X-ray information to binocular visual information was completed.

Aiming at the extraction of gangue grasping features, a method of generating gangue optimal grasping cube by point cloud processing is proposed. The pre-processing method which can filter out the surrounding environment and background of site cloud information is analyzed. The interference problem of site cloud information to target gangue point cloud is solved by using direct filter. The point cloud of target gangue was reduced by voxel filter, the amount of point cloud data was reduced, and the efficiency of subsequent point cloud processing was improved. The radius filter with stronger robustness is used to solve the interference problem of point cloud noise. Different point cloud segmentation methods are compared and analyzed, and DBSCAN clustering algorithm is used to segment target gangue point cloud data. The minimum direction bounding box OBB algorithm is used to construct the optimal grasping cube of gangue, and the characteristic parameters of gangue grasping are extracted.

Finally, the above proposed extraction method of gangue grab feature combining X-ray image and binocular visual information was verified experimentally, and the whole experimental scheme was designed. The static experiment, dynamic experiment and gangue grab feature extraction experiment were carried out. The results showed that, In this paper, the target gangue grasping feature extraction method is to prepare gangue grasping feature parameters in advance when the manipulator snatches gangue, which can obviously improve the success rate and efficiency of the manipulator snatching gangue.