Scanning Bot

Abstract

Panoramic RGB-D cameras enable high-quality 3D scene reconstruction but require manual viewpoint selection and physical camera transportation, making the process time-consuming and tedious-especially for novice users. Key challenges include ensuring sufficient feature overlap between camera views and planning collision-free paths. We propose a fully autonomous scan planner that generates efficient and collision-free tours with adequate viewpoint overlap to address these issues. Experiments in both synthetic and real-world environments show that our method achieves up to 99% scan coverage and is up to three times faster than state-of-the-art view planning approaches.

Our approach

Scanning bot pipeline: The set-covering procedure first identifies the optimal viewpoints. This is followed by TSP-based, collision-free view planning to autonomously scan a scene.

We generate 2d grid map by exploration: Autoexplorer. Then, we select viewpoints with greedy set covering-based approach.

Viewpoint Selection

Our algorithm efficiently selects the best viewpoints to cover a region. It first identifies the boundary cells of each region using DFFP from Neuro-Explorer, then evaluates candidate points on the boundary with a cost function that balances the relative size of the region and the distance from nearby obstacles:

$\phi(x) = \frac{|S_i|}{\max_i |S_i|} - e^{-\lVert x - \mathbf{o} \rVert_2}$

The point with the highest score is chosen as the next viewpoint, and this process is repeated until all target areas are fully covered.

Scan Order Planning

We begin with a TSP solver to create an initial scan plan, but if some paths are infeasible, our method searches for detours using visibility and roadmap graphs. To choose the better route, we balance travel distance with the extra time needed for new viewpoints using the cost function:

$\psi(\mathcal{\tau}) = (1-\eta)\sum_{i}^{m-1}\rho(x_{i+1},x_i) + \eta \cdot n$

This ensures the final scan sequence is both efficient and feasible, with global paths planned by A* and local motion refined by the TEB planner.

Experimental Results

Performance Analysis in Synthetic Environments

Qualitative comparisons of the sim results: Top row shows the selected viewpoints (red dots) and the covered area (yellow pixels). The bottom row shows the planned path (green lines).

[Table 1] Quantitative comparisons of the sim results: This table shows coverage and number of viewpoints for different methods in various simulation worlds. "-" indicates that no feasible plan was generated within a finite time.

[Table 2] Quantitative comparisons of the sim results: The table shows exploration and planning times with baseline methods in sim environments. The exploration time is identical for BCD, CLCPP, GKVM, and our method, as they all use the map generated by Autoexplorer.

Overall, our method generates the fewest viewpoints while achieving over 99% scanning coverage; see [Table 2] and [Table 3] This is because our method incorporates the visibility constraint in view plan optimization, whereas other approaches do not consider such a constraint. The two CPP-based methods achieved faster planning times than our approach. However, the planning time constitutes a relatively small portion of the total scanning time, which includes planning, navigation, and image capturing. This is because the number of viewpoints primarily influences the total scanning time.

Real World Experiments

Qualitative comparisons of real-world experiments: The top row shows viewpoints (red) and covered areas (yellow). The bottom row shows 3D models with detailed textures, highlighting our method's better-textured mesh quality.

**[Table 3] Comparisons of the real-world results**

The experiments were conducted in the ASAN Engineering Building at Ewha Womans University, which consists of a long corridor. [Table 3] presents the quantitative results of these experiments. We observed that our method and CLCPP achieved over 99% coverage of the scanning space. However, our method was approximately three times faster than CLCPP in terms of total scanning time. The total scanning time encompasses visiting all planned viewpoints and capturing images at each point.

BibTeX


        @inproceedings{Leehankim2025_iros,
        author = {Eujeong Lee*, Kyung Min Han*, and Young J. Kim},
        title = {Scanning Bot: Efficient Scan Planning using Panoramic Cameras},
        year = {2025},
        pages = {1--7},
        booktitle = {Proc. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
        }