Research On Object 6D Pose Estimation Based On RGB-D Data

6D pose estimation of object is a key task for understanding the given scene,However,accurate 6D pose estimation is still a challenge problem due to the complexity of application scenarios caused by illumination changes,occlusion and even truncation between objects.The traditional pose estimation method achieves pose estimation by first extracting manual features of the target object and then establishing corresponding relations or matching template.However,this method is only suitable for objects with rich texture and has poor robustness.In view of the recent success of deep learning in visual recognition,a series of neural networks based on deep learning are introduced to estimate the object 6D pose.Compared with traditional methods,these methods can have better occlusion resistance,less consuming time and higher accuracy.Therefore,for 6D object pose estimation in the complex scenes,this paper designs a novel attentive multi-scale pose estimation network and an instance segmentation module based on instance center clustering.The main research contents are as follows:(1)The research meaning and background of 6D pose estimation is introduced in the Chapter 1.An in-depth investigation on 6D pose estimation algorithms based on deep learning is conducted in Chapter 2.Existing methods are divided into methods based on RGB data and methods based on RGB-D data according to input data types,and their features and datasets are analyzed and discussed.(2)With RGB-D data as input,an attentive multi-scale pose estimation network is designed in Chapter 3.The pixel-level feature attention mechanism is used to efficiently extract and fuse the color features and geometric features in the input data,and a multi-scale network is used to extract and utilize the contextual information of the target object.The network can achieve efficient and accurate object 6D pose estimation,and exhibits state-of-the-art performance on public Linemod and YCB-Video datasets.(3)Aiming at the problem that the existing pose estimation network performs poorly in instances with similar appearances and different sizes,an instance segmentation module based on instance center clustering is proposed in Chapter4.Accurate instance segmentation can provide correct target object information for pose estimation,so the accuracy of object 6D pose estimation is improved.Experimental results show that the addition of the instance segmentation module greatly improves the accuracy of the pose estimation network on the Linemod and YCB-Video datasets.(4)The pose estimation network can be applied in the actual production and life scenarios.A robotic grasping system is built experimentally in Chapter 5.The system includes the manipulator,RGB-D camera and end effector.The camera calibration and hand-eye calibration are performed so that the camera coordinate system and the manipulator coordinate system can be converted to each other,which means that the manipulator can use the object 6D pose output by the pose estimation network for the grasping task.