Manipulator grasping system based on multi depth camera fusion
To solve the problems of noise interference and occlusion in the object depth image obtained from the per-spective of single depth camera,the shape of object point cloud is missing or deformed,and the three-dimensional features are unstable.Therefore,a multi depth camera fusion method was proposed.Using the results of multi-camera calibration,the point cloud data from various camera perspectives were stitched together.Noise points in the point cloud were removed through k neighborhood denoising algorithm.Subsequently,the denoised point cloud data underwent voxelization and subsampling methods to ensure a uniform spatial distribution of the stitched point cloud,thereby obtaining a complete representation of the object's shape.Simultaneously,an end-to-end neural network was established to predict the 3D position and 3D orientation of a two-finger parallel gripper based on RGB-D images cap-tured by multiple cameras.To optimize the predicted grasping results,the robustness evaluation of the prediction results in neural network was improved,and the evaluation method of force spiral space field was proposed,which improved the success rate by 4％.Using AR5 manipulator in the actual grasping system,the grasping success rate was 91％.
NETL
NSTL
万方数据
