HeadScanning

This script is a complete Python pipeline for creating 3D models using an Intel RealSense D435 mounted on a Universal Robots UR5e. The script captures point cloud data, transforms captured data into a common coordinate frame, statistically denoises the merged point cloud using MCMD_Z, fills the base of the cleaned, merged cloud, and performs Poisson Surface Reconstruction to output a watertight STL model with a flat base. This pipeline was developed as part of a Master's thesis in Applied Mathematics and Statistics at Cal Poly Pomona.

Example output:

Project Structure

HeadScanning/
├── head_scanner.py     # Main pipeline
├── requirements.txt    # Python dependencies
└── README.md

---

Setup Guide

This guide assumes no prior experience. Follow each step carefully.

Step 1 — Hardware setup

Before running anything, make sure you have:

• A Universal Robots UR5e powered on, in Remote Control mode, connected to the local network
• An Intel RealSense Depth Camera D435 mounted directly to the robot's tool flange.
• The camera connected to your computer via USB-C
• A flat surface in front of the robot where the scanned object will be placed

Step 2 — Install Python

You need Python 3.10 or newer.

• Check if you already have it:

  python3 --version
  

  If this prints Python 3.10.x or higher, skip to Step 3.

• Install Python (if needed):

  • Go to https://www.python.org/downloads/
  • Download the latest stable release and run the installer
  • On the Windows installer, check "Add Python to PATH" before clicking Install

Step 3 — Get the Code

Open a terminal (Mac/Linux) or Command Prompt (Windows) and run:

git clone https://github.com/blainebehen/HeadScanning.git
cd HeadScanning

Don't have git? Download it from https://git-scm.com/downloads, or download the project as a ZIP from GitHub and unzip it.

Step 4 — Install Dependencies

pip install -r requirements.txt

Optional: if you want a GUI to verify the camera is working before running the pipeline, install Intel RealSense Viewer from https://github.com/realsenseai/librealsense/releases. Make sure to close RealSense Viewer before running head_scanner.py — only one program can access the camera at a time.

Step 5 — Configure the Script

Open head_scanner.py in a text editor and edit the USER SETTINGS block near the top:

• ROBOT_IP — the IP address of the UR5e on the local network
• OUT_DIR — the directory where output files will be saved

The other settings (motion speed, bounding box, MCMD parameters, Poisson parameters) have working defaults.

---

Running the Pipeline

With the robot powered on, in Remote Control mode, and the camera connected:

python head_scanner.py

The script will prompt you for an object name, then automatically:

1. Drive the robot through the seven capture poses
2. Capture a 3D point cloud at each pose
3. Merge the point clouds into the robot's base frame
4. Denoise the merged point cloud with MCMD_Z
5. Fill the underside of the object with synthetic base points
6. Mesh the result with Poisson Surface Reconstruction
7. Slice the bottom flat and export the final STL

When it finishes, you'll have these files in your OUT_DIR (filenames use the sanitized object name — mug shown here as an example):

File	Description
merged_pointcloud_mug.ply	Raw merged point cloud, all seven captures combined
MCMD_cleaned_pointcloud_mug.ply	Merged point cloud after MCMD_Z denoising
MCMD_cleaned_pointcloud_filled_mug.ply	Denoised cloud with synthetic base added
poisson_mesh_mug.stl	Watertight mesh from Poisson reconstruction
flat_base_mug.stl	Final mesh with flat base — this is the one you print

Per-pose data (color_0.png, depth_0.png, cloud_pos_0.ply, etc.) is also saved for debugging.

---

Coverage Region

The seven-pose protocol is calibrated for objects placed on a flat surface in front of the robot at approximately (-41, -400, -17) mm relative to the robot base.

For best results, the object should fit within these bounds:

• x ∈ [-173, 95] mm
• y ∈ [-550, -264] mm
• z ∈ [-17, 220] mm

Objects outside this region may not be fully captured by the seven pose protocol.

---

Tips

• Matte, opaque objects scan best. Clear plastics and shiny metals can cause sparse point cloud sampling or meshing artifacts.
• Fine details smaller than ~7 mm get smoothed. From the Poisson reconstruction at the default octree depth. Increase POISSON_DEPTH to preserve them, at the cost of more noise.
• Don't move the object or table between scans. The scanning protocol is designed for a fixed setup.
• For specular (shiny) objects: oblique camera angles capture data better than direct ones — see Section 3.2 of the thesis.

---

Citation

If you use this code in academic work, please cite the thesis:

Behen, B. (2026). Building a 3D Scanner. Master's thesis, California State Polytechnic University, Pomona.

Acknowledgments

Developed under the supervision of Dr. Arlo Caine in the Department of Mathematics and Statistics at Cal Poly Pomona.