Abstract:Three-dimensional(3D) environmental reconstruction of the manipulator workspace is a prerequisite to enable autonomous collision-free path planning and subsequent task execution, wherein the positioning accuracy of the equipped sensor is one key issue to determine the performance of 3D reconstruction. In this paper, an eye-in-hand RGB-D camera is used to complete the manipulator workspace scanning and a novel high-accuracy pose estimation method for 3D reconstruction is proposed, which not only considers the geometric and photometric constraints constructed from the visual information used in traditional methods, but also considers the manipulator motion constraints since the positioning accuracy of the manipulator is usually high. Based on above three constraints, three error functions are constructed respectively and fused with different weights to form an integrated objective function. Then, the problem is converted into solving the least squares problem by the Gaussian Newton nonlinear iterative method, which involves the computation of Jacobian matrices and residuals of the three error functions. To this end, the mathematical derivation of their formulas considering the manipulator motion constraints are given in detail. By qualitative and quantitative experimental comparison with traditional methods, the proposed method can greatly improve the accuracy of pose estimation and 3D reconstruction, which always maintains the error within 3 cm, and show good robustness even in the challenging environments with weak texture or geometric features. (Supplemented video link: https://youtu.be/FkFTxssnWms)

Volumetric View Planning for 3D Reconstruction with Multiple Manipulators

An Efficient Projection-Based Next-best-view Planning Framework for Reconstruction of Unknown Objects

Volumetric Information Gain Guided Receding Horizon Planner for Active 3D Reconstruction.

Bag of Views: An Appearance-Based Approach to Next-Best-View Planning for 3D Reconstruction

MAP-NBV: Multi-agent Prediction-guided Next-Best-View Planning for Active 3D Object Reconstruction

PC-NBV - A Point Cloud Based Deep Network for Efficient Next Best View Planning.

A Global Max-Flow-Based Multi-Resolution Next-Best-View Method for Reconstruction of 3D Unknown Objects.

Volumetric Next Best View by 3D Occupancy Mapping Using Markov Chain Gibbs Sampler for Precise Manufacturing

3d Reconstruction With Multi-View Texture Mapping

An efficient planning method for mobile robot based 3D digitization

Viewpoint planning for automated 3D digitization using a low-cost mobile platform

Visual Odometry Based 3D-Reconstruction

Model-Based Active Viewpoint Transfer For Purposive Perception

NBV/NBC Planning Considering Confidence Obtained From Shape Completion Learning

High-Accuracy Pose Estimation for 3D Reconstruction Considering the Manipulator Motion Constraints with an Eye-in-Hand RGB-D Camera

NBV-SC: Next Best View Planning based on Shape Completion for Fruit Mapping and Reconstruction

An Automated View Planning Method of Robot 3D Measurement.

Sparse sampling-based view planning for complex geometries

Embedding Visual Cognition in 3D Reconstruction from Multi-View Engineering Drawings

Next Best Viewpoint (NBV) Planning for Active Object Modeling Based on a Learning-by-Showing Approach.

SCVP: Learning One-Shot View Planning via Set Covering for Unknown Object Reconstruction