| 摘要: |
| 目的 现有的物体六自由度姿态估计方法侧重于处理已训练过的对象,针对未知物体纹理细节较弱或在有
遮挡、光照复杂的非结构化环境中的六自由度姿态估计仍是一个具有挑战性的问题。 方法 提出了一个物体六自由
度姿态估计的网络,首先通过局部特征匹配获取图像对中足够多的匹配点对;其次,通过传感器读取匹配点对的深
度信息得到相应点云数据并将其作为后续点云精准配准的初始点云进行配准,最终得到源点云相对于目标点云的
旋转矩阵和平移向量,即未知物体在机器人坐标系下的六自由度姿态。 结果 该网络在基准数据集上估计的 6D 姿
态结果较好,dADD-S 值为 80. 0%;在 Occlusion Linemod 数据集上 dADD-S 值也达到了 78. 0%,均表现出了非常优异的
性能。 结论 该网络泛化性比较好,不仅能够准确地估计严重遮挡、背景杂波和光照差等条件下物体的六自由度姿
态,而且对随机噪声也具有较好的鲁棒性。 |
| 关键词: 姿态估计 局部特征匹配 点云配准 未知物体 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Pose Estimation Network Combining Local Feature Matching and Point Cloud Registration |
|
ZHANG Jing1,YU Ling2
|
|
1. School of Computer and Information Anhui Polytechnic University Wuhu 241000 Anhui China
2. Yangtze River Delta Hart Robot Industry Technology Research Institute Wuhu 241000 Anhui China
|
| Abstract: |
| Objective Existing six-degree-of-freedom 6-DoF object pose estimation methods mainly focus on dealing with
trained objects. Estimating the 6-DoF pose of unknown objects with weak texture details or in unstructured environments
with occlusion and complex lighting remains a challenging problem. Methods A novel network for 6-DoF object pose
estimation was proposed. First a sufficient number of matching point pairs were obtained from image pairs through local
feature matching. Second the depth information of these matching point pairs was read via sensors to generate the
corresponding point cloud data which was used as the initial point cloud for subsequent high-precision point cloud
registration. Finally through the registration process the rotation matrix and translation vector of the source point cloud
relative to the target point cloud were obtained. The matrix and vector represented the 6-DoF pose of the unknown object
in the robot coordinate system. Results The proposed network demonstrated strong performance in 6D pose estimation on
benchmark datasets with a dADD-S
value of 80. 0%. It also achieved a dADD-S
value of 78. 0% on the Occlusion Linemod
dataset demonstrating outstanding performance. Conclusion This network exhibits good generalization ability. It can
accurately estimate the 6-DoF pose of objects under conditions such as severe occlusion background clutter and poor
lighting. Moreover it shows good robustness against random noise. |
| Key words: pose estimation local feature matching point cloud registration unknown object |
