【摘要】 目标6D位姿检测是实现机器人自主抓取的关键。为克服传统点对识别（PPF）方法检测性能差、耗时及难以检测到多平面特征工件的6D位姿等不足，提出面向无序分拣场景的工件6D位姿检测方法。首先，基于模型平面点分布筛选多平面特征工件，提取其边界特征进行6D位姿检测，并在多视点下提取模型点对以去除冗余点对，提高算法识别速度。其次，匹配场景与模型间的点对特征，利用快速投票方案获取无序场景中目标的位姿假设集合。接下来，通过位姿验证筛选方法，剔除重复和误匹配位姿，实现目标多实例位姿的粗略估计，并借助迭代最近点（ICP）算法完成目标位姿的精确估计。实验结果表明：在无序仿真场景中，单次识别时间小于等于1.15 s，平均平移偏差小于等于0.95 mm，平均旋转误差小于等于1.56°；在实际场景中，平均识别成功率为95.82%，平均单次识别时间为1.11 s。综上，该6D位姿检测方法在保证识别效率的同时兼顾了位姿估计精度，并在识别精度和速度上均优于同类算法，为机器人的精准抓取的实现提供了有力的保障。更多还原

【Abstract】 Objective 6D pose detection is a key technology for enabling autonomous grasping in robots. Currently, traditional point-pair feature(PPF) methods face three major challenges: 1) excessive sensitivity to sensor noise, severe occlusions, and background clutter; 2) reduced matching performance when the workpiece has numerous repetitive features; 3) slow recognition speed due to the need to search many point pairs and compute transformation relationships. This study proposes a point-pair feature-based 6D pose detection method designed for robotic sorting system grasping tasks.Methods Firstly, multi-plane feature workpieces were screened based on distributions of model plane points, and their boundary features were extracted for 6D pose detection. Model point pairs were extracted from multi-view points to remove redundant point pairs and improve the recognition speed of algorithms. Secondly, to further enhance recognition speed, a method was employed to extract model point pairs from multiple viewpoints, which helped in eliminating redundant point pairs that did not contribute to the detection process. Thirdly, the point-to-point characteristics between scenes and models were matched, and a fast voting scheme was employed to obtain pose hypothesis sets for targets in a disordered scene. Then, a pose verification and screening method was introduced to eliminate duplicate and mismatched poses, which was essential for realizing a rough estimation of multi-instance poses for the targeted workpieces. Finally, an algorithm called Iterative Closest Points(ICPs) was utilized to refine the rough estimates and achieve a more accurate estimation of the targeted poses.Results and Discussions Experimental results showed that in the context of disordered simulation scenes, the proposed method demonstrated a single recognition time of ≤ 1.2 seconds, with an average translation deviation of ≤ 1 mm and an average rotation error of ≤ 1.56°. These results indicated a high level of precision and efficiency in pose detection. In an actual scenario, this method achieved an average recognition success rate of 95.8%, with an average single recognition time of 1.1 seconds. The high success rate and rapid speed highlighted that this method has favorable practical applicability in robotic sorting tasks. Therefore, this study highlighted that the proposed 6D pose detection method significantly outperformed the original PPF algorithm in terms of recognition speed, while also improving the accuracy of pose estimation. This advancement was crucial for the reliable and efficient operation of robotic systems in precision grasping applications. Finally, this 6D pose detection method not only ensured recognition efficiency but also accounted for the accuracy of pose estimation. This meant that recognition speed was significantly improved compared to the original PPF algorithm, providing a strong guarantee for the realization of accurate robotic grasping.Conclusions The research presents a comprehensive approach to enhancing 6D pose detection in disordered sorting scenarios, representing a significant advancement in robotic vision and grasping technologies. The proposed method is verified as effective in both simulated environments and real-world working conditions. In addition, it demonstrates superior performance compared to existing approaches, supporting more accurate analysis in 6D pose detection applications.更多还原