PB-NBV: Efficient Projection-Based Next-Best-View Planning Framework for Reconstruction of Unknown Objects

TL;DR

This work tackles next-best-view planning for complete 3D reconstruction of unknown objects, where ray-casting in voxel representations creates a bottleneck. It introduces a projection-based NBV framework that fits voxel-occupied and frontier regions into ellipsoids via a Gaussian Mixture Model and selects viewpoints by projecting these ellipsoids into the camera view, effectively replacing ray-casting. A global partitioning strategy is added to stabilize registration and reduce backtracking, leading to improved convergence speed and higher coverage. Evaluations in simulation and real-world experiments show the approach achieves higher coverage with lower computational cost than several baselines, and the authors plan to release open-source software.

Abstract

Completely capturing the three-dimensional (3D) data of an object is essential in industrial and robotic applications. The task of next-best-view (NBV) planning is to calculate the next optimal viewpoint based on the current data, gradually achieving a complete 3D reconstruction of the object. However, many existing NBV planning algorithms incur heavy computational costs due to the extensive use of ray-casting. Specifically, this framework refits different types of voxel clusters into ellipsoids based on the voxel structure. Then, the next optimal viewpoint is selected from the candidate views using a projection-based viewpoint quality evaluation function in conjunction with a global partitioning strategy. This process replaces extensive ray-casting, significantly improving the computational efficiency. Comparison experiments in the simulation environment show that our framework achieves the highest point cloud coverage with low computational time compared to other frameworks. The real-world experiments also confirm the efficiency and feasibility of the framework. Our method will be made open source to benefit the community.

Figures (12)

• Figure 1: Overview of the NBV planning experiment platform. An object to be measured is placed on a turntable, and a 3D camera is equipped at the end of the robotic arm for data acquisition. Our NBV planning framework accomplishes a complete reconstruction of the object by controlling the robotic arm and the turntable.
• Figure 2: Overview of NBV planning framework. The orange arrows describe the running steps of the NBV iteration process. After capturing the point cloud, the framework performs preprocessing and pose registration. The registered point cloud is then input into the ellipsoid representation module, where it is converted into a voxel structure and fitted into an ellipsoid. The projection evaluation function is used to assess all candidate viewpoints, and a global partitioning strategy selects the optimal viewpoint for the next frame. Finally, the robotic arm moves to the selected viewpoint and begins the next iteration.
• Figure 3: The components of the radius of the candidate viewpoint sampling region and the results of candidate viewpoint sampling within this region.
• Figure 4: Use a 2D grid to describe the voxel classification. (a) Classification results of a single frame input. (b) Classification results of multiple frames input. The green box represents the object's bounding box.
• Figure 5: Results of different GMM clustering numbers. $T_o$ represents the number of clusters of occupied voxels, and $T_f$ represents the number of clusters of frontier voxels.