A method for estimating an unknown target grasping pose based on keypoint detection

Abstract

Aiming at the difficulty of estimating the unknown target grasping pose in robotic applications, a pose estimation method based on crucial point detection is proposed, and the CenterNet-SPF (Subpixel- Feature Pyramid Networks) unknown object grasping pose estimation network is designed. CenterNet-SPF divides the grasping angle into 18 feature point types, sorts, and filters the weights of the feature points to obtain the key point category and corresponding point coordinates and then regresses the key points to get the grasping frame parameters, completing the target detection network to the grasping detection network. Improve the generalization ability of unknown target detection; design a feature extraction network structure for dual-channel feature fusion, introduce sub-pixel convolution instead of transposed convolution to perform dual-channel feature fusion for high and low-level features and increase the sampling convolutional layers, increases the feature output resolution to improve prediction accuracy. The CenterNet-SPF grasping pose estimation network was verified on the Cornell grasping dataset, and the prediction accuracy reached 98.31%; the generalization ability was tested on the untrained Jacquard dataset, and the recognition accuracy reached 95.5%; the actual application tests the success rate of middle grasping is 88.7%. The experimental results show that the proposed network can accurately detect the grasping pose of unknown targets, has good generalization ability, and can meet the application requirements of grasping tasks.