6D Pose Estimation Of Target Object For Robot Grasping

6D pose estimation is a key technology in the field of computer vision,which can obtain complete spatial information of the estimated target.It is widely used in intelligent robot operation,aerospace target recognition,target tracking and other modern scientific and technological fields.Aiming at the accuracy of target pose estimation and the adaptability of complex environment,this paper conducts in-depth research and theoretical innovation on pose estimation technology from two aspects of traditional image features and deep learning,proposes an adaptive multineighborhood feature point algorithm for weak texture images and a voting network that minimizes distance loss,and verifies the superiority of the algorithm through multi-scene experiments.The main research contents are as follows:(1)Aiming at the problems that the existing feature point algorithms can’t extract effective descriptors in the face of weak texture images and the accuracy of feature matching is low,a new descriptor(AMST)suitable for various texture images is proposed based on the structure information of image.Based on a series of image neighborhoods and their structure tensors,the algorithm can express the texture information of images in multiple layers and solve the problem of feature extraction and matching in weak texture scenes.At the same time,the density of feature points adapt to the number of neighborhood,and the Hessian matrix is used to screen feature points to improve speed and stability of the algorithm.(2)Aiming at the problems of low accuracy of existing 6D estimation networks and poor adaptability to complex scenes,an efficient 6D estimation neural network(MDL)is built which can effectively deal with complex problems such as occlusion and truncation.MDL is designed as a voting network to maximize the prediction of invisible key points.The whole network adopts residual structure to improve the problem of gradient disappearing and gradient dispersion.FPS algorithm is used to select key points to enhance the adaptability of the algorithm to different backgrounds.Based on the loss of direction,the MDL introduces distance loss to reduce the influence of the distance between pixels and key points on voting and improve the prediction accuracy.(3)Complete the 6D pose estimation based on MDL and AMST,The aim of the algorithms is to predict the key points distributed on the surface of the target object,and then solve the PnP problem to obtain the 6D pose of the target.In this paper,EPnP algorithm,an improved version of PnP,is used to obtain the 6D pose.The core idea of EPnP is to represent the three-dimensional coordinate points as a combination of four control points and optimize only for the four control points,which is very fast.And the solution takes into account at most four singular vectors,so the precision is also very high.(4)Based on MDL and AMST,a complete grasping system is built on the hardware basis of AUBO-i5 robotic arm and zed camera to verify the accuracy and effectiveness of the algorithms in practical application.On the basis of the above theoretical research,a large number of experiments were carried out in this paper,including AMST multi-texture image experiment,AMST and MDL algorithm comparison experiment with other classical algorithms,AMST and MDL comparison experiment,pose estimation experiment and grab experiment,etc.The accuracy and adaptability of the algorithm in complex environment is verified by these experiments.