基于三维点云处理技术的工件识别和匹配

Parts Identification and Matching Based on 3D Point Cloud Processing Technology

摘要: 在自动分拣领域，机器视觉技术起到举足轻重的作用，本文采用一种基于三维点云处理技术的方法对工件进行识别和匹配，从而从生产线上的杂乱工件堆中分拣出目标工件。首先，利用激光传感器采集密集点云，得到识别区域的三维点云数据。其次，进行点云数据的去噪和过滤，对去噪后的点云计算法线向量，再进行体素格子下采样获得关键点，减少计算量。再次，对关键点计算法线方向的直方图特征描述子，从而获取目标点云和场景点云相关联的描述符，使用K近邻树状结构搜索算法，基于欧几里德距离找到相似的描述符，并添加到关联集中。再者，调用聚类算法，基于霍夫投票算法获得场景中的与目标点云匹配的局部点云集，然后使用点云匹配算法迭代最近点算法进行目标点云与锁定的局部点云集进行匹配。最后，为了减少错误率，使用全局假设验证算法过滤掉错误的点，并标定识别到的目标进行可视化输出。该方法相对于一般算法，实现对目标的高效匹配，提高了识别的时间和精度。

Abstract: In the field of automatic sorting, machine vision technology plays a decisive role. This paper uses a method based on 3D point cloud processing technology to identify and match workpieces, so as to sort out the target workpieces from the messy workpiece piles on the production line. Firstly, a dense point cloud is collected by using a laser sensor to obtain three-dimensional point cloud data of the identified area. Secondly, after denoising and filtering the point cloud data, it calculates the normal vector of the denoised point cloud, and then performing the voxelgriddownsampling to obtain key points and reduce the calculation amount. Thirdly, the histogram feature descriptor of the normal direction is calculated for the key point, thereby it obtains the descriptor associated with point cloud of the target and point cloud of the scene. It uses the K-tree structure search algorithm to find a similar descriptor based on the Euclidean distance, and added to the association set. Fourthly, the clustering algorithm is called to obtain a local point cloud set matching the target point cloud in the scene based on the Hough voting algorithm, and then the target point cloud is matched with the locked local point cloud set using the classical point cloud matching algorithm ICP. Finally, the global hypothesis verification algorithm is used to filter out the wrong points to reduce the error rate and to calibrate the identified targets for visual output. Compared with the general algorithm, the method is more efficient and improves the time and accuracy of recognition.