基于深度学习的物体点云六维位姿估计方法  

Deep Learning-Based 6D Object Pose Estimation Method from Point Clouds

作　　者：李少飞 史泽林 庄春刚[1] LI Shaofei;SHI Zelin;ZHUANG Chungang(School of Mechanical Engineering,Shanghai Jiao Tong University,Shanghai 200240,China)

机构地区：[1]上海交通大学机械与动力工程学院,上海200240

出　　处：《计算机工程》2021年第8期216-223,共8页Computer Engineering

基　　金：国家自然科学基金(51775344)。

摘　　要：物体位姿估计是机器人在散乱环境中实现三维物体拾取的关键技术,然而目前多数用于物体位姿估计的深度学习方法严重依赖场景的RGB信息,从而限制了其应用范围。提出基于深度学习的六维位姿估计方法,在物理仿真环境下生成针对工业零件的数据集,将三维点云映射到二维平面生成深度特征图和法线特征图,并使用特征融合网络对散乱场景中的工业零件进行六维位姿估计。在仿真数据集和真实数据集上的实验结果表明,该方法相比传统点云位姿估计方法准确率更高、计算时间更短,且对于疏密程度不一致的点云以及噪声均具有更强的鲁棒性。Object pose estimation is a key technology required for enabling the robots to pick 3D objects in a cluttered environment.However,most of the existing deep learning methods for pose estimation rely heavily on the RGB information of the scene,which limits their applications.To address the problem,a deep learning-based method for 6D object pose estimation is proposed.A data set for industrial parts is generated from physical simulation,and then the 3D point cloud is mapped to the 2D plane to generate a deep feature map and normal feature map.On this basis,a feature fusion network is used for 6D pose estimation of industrial parts in cluttered environments.Experimental results on the simulation data set and the real data set show that the proposed method improves the accuracy of pose estimation and reduces time consumption compared with traditional point cloud pose estimation methods.In addition,the method displays high robustness to the point clouds with different density and noises.

关 键 词：点云 位姿估计 特征融合 深度学习 损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]