基于改进YOLOv7和点云配准的零件识别与位姿估计方法  

Part Recognition and Position Estimation Method Based on Improved YOLOv7 and Point Cloud Alignment

作　　者：张钊 王吉芳[1] 刘相权[1] 郭凯旋 王凯 ZHANG Zhao;WANG Jifang;LIU Xiangquan;GUO Kaixuan;WANG Kai(College of Mechanical and Electrical Engineering,Beijing Information Science&Technology University,Beijing 100192,China)

机构地区：[1]北京信息科技大学机电工程学院,北京100192

出　　处：《机床与液压》2025年第5期81-87,共7页Machine Tool & Hydraulics

摘　　要：针对工业环境中常见的弱纹理堆叠零件的识别与位姿估计挑战,提出一种融合YOLOv7目标检测算法和点云配准技术的综合方法。首先,利用改进的YOLOv7算法快速识别并定位零件的二维位置,随后将包含零件的二维区域对齐深度图转换为对应的三维点云。在点云处理阶段,采用深度阈值分割和欧氏聚类分割方法以分离目标零件与背景及其他干扰物体。位姿估计环节采用SAC-IA算法进行粗配准,接着通过引入自适应权重机制和全局优化策略的改进ICP算法实现精配准,以获得零件的最终6D位姿。该改进策略显著优化了点对选择和配准流程,增强了算法的鲁棒性和准确性。在公开的零件数据集上进行实验验证,结果表明:提出的零件识别与位姿估计方法能够实现对不同形状、弱纹理、散乱堆叠零件的快速、准确识别和位姿估计,其中位置误差在1 mm以内,角度误差在1°以内,满足实际应用需求,展示了其在工业自动化领域的实用性和有效性。To address the challenges of recognizing and estimating the position of untextured stacked parts,which are common in industrial environments,an integrated approach that incorporated the YOLOv7 target detection algorithm and point cloud alignment techniques was proposed.The improved YOLOv7 algorithm was utilized to quickly identify and locate the 2D position of the part,followed by converting the aligned depth map of the 2D region containing the part into a corresponding 3D point cloud.In the point cloud processing stage,depth threshold segmentation and Euclidean clustering segmentation methods were used to separate the target part from the background and other interfering objects.In the pose estimation stage,the SAC-IA algorithm was used for coarse alignment,followed by an improved ICP algorithm that introduced an adaptive weighting mechanism and a global optimization strategy for fine alignment to obtain the final 6D pose of the part.The improved strategy significantly optimizes the point-pair selection and alignment process and enhances the robustness and accuracy of the algorithm.Through experimental validation on publicly available part-piece datasets,the results show that using the part recognition and position estimation method proposed in this study,fast and accurate recognition and position estimation for different shapes,non-textured,and scattered stacked parts can be realized,while the position error is within 1 mm and the angular error is within 1°,which meets the requirements of practical applications and demonstrates its practicability and effectiveness in the field of industrial automation.

关 键 词：弱纹理堆叠零件 零件识别 位姿估计 深度学习 

分 类 号：TP242.2[自动化与计算机技术—检测技术与自动化装置]