This paper, titled "Development of an Automatic 3D Human Head Scanning-Printing System," presents a complete end-to-end pipeline for creating physical 3D-printed busts of human subjects using consumer-grade hardware and a mobile robot.
TL;DR
The researchers developed a fully automated system that uses a robot-mounted Kinect v2 and an IMU suite to scan a human head and produce a 3D-printed model. By combining hardware-based motion tracking with software-based reconstruction, the system achieves professional-grade results at a fraction of the cost of commercial laser scanners, all while eliminating the need for manual 3D model "cleaning."
What Problem Does This Paper Address?
High-quality 3D head models are essential for fields like anthropological studies, digital avatar creation, and ergonomic product design (e.g., fitting helmets or glasses). However, the industry faces a "Goldilocks" problem:
- Professional Scanners: Systems like Cyberware or FastSCAN are highly accurate but prohibitively expensive and often require subjects to remain perfectly still for long periods.
- Consumer Scanners: Low-cost sensors like the Microsoft Kinect are accessible but prone to "drift" (accumulative errors) and usually produce "noisy" raw data that includes the background, floor, or table.
Before this paper, turning a raw scan into a 3D-printable file required significant manual labor to trim away the environment and scale the model correctly. The authors sought to automate this entire "scan-to-print" workflow.
Methods
The authors built a custom hardware stack and a multi-stage software pipeline:
- The Hardware: A Microsoft Kinect v2 (depth sensor) was bundled with a "composite sensor" (gyroscope, accelerometer, and digital compass). This entire rig was mounted on a DFRobot HCR Mobile Robot, which was programmed to circle the subject at a 1-meter radius.
- Sensor Fusion: To solve the "drift" problem, the system uses a Kalman filter to fuse IMU data with the Kinect’s depth stream. When estimating the camera's position (pose estimation), they assigned a 0.8 weight to the software-based ICP (Iterative Closest Point) algorithm and a 0.2 weight to the hardware sensors. This "hardware-assisted" approach keeps the reconstruction stable even if the software loses track of the subject's features.
- Automatic Selection (Segmentation): This is the paper's most practical contribution. They developed two algorithms to automatically crop the head from the environment:
- For inanimate models: It uses RANSAC plane fitting to identify the table and crops everything above it.
- For real humans: It calculates the 3D centroid of the robot’s path, finds the "nearest neighbors" to identify the top of the head, and creates a virtual bounding prism to extract the bust.
- Reconstruction: The system uses a modified version of KinectFusion to create a Truncated Signed Distance Function (TSDF) volume, which is then converted into a polygon mesh using the Marching Cubes algorithm.
Datasets
The authors evaluated their system using:
- A standard inanimate human head model (for ground-truth comparison).
- Four real human subjects (two male, two female).
- Comparative Benchmarks: They compared their results against scans from the FastSCAN and Cyberware commercial laser systems.
- The authors have published their data on The MCRLab 3D Human Head Scanning Repository.
Key Insights and Ideas
- Hardware as a Safety Net: Purely vision-based reconstruction (like standard KinectFusion) often fails on smooth or repetitive surfaces. By adding a $20 IMU suite, the authors provided a "ground truth" for rotation and acceleration that prevents the software from getting "lost."
- Hair is a Laser's Weakness: Interestingly, the authors found that their low-cost system sometimes outperformed expensive laser scanners when dealing with human hair. Lasers often struggle with "optically uncooperative" materials like dark or shiny hair, whereas the Kinect’s Time-of-Flight (ToF) approach proved more robust in capturing hairstyles and ponytails.
- The "Ready-to-Print" Goal: Most research stops at the 3D mesh. This paper emphasizes the post-processing—automatic scaling and "virtual plane" selection—necessary to make the system usable by non-experts.
Key Results
- Accuracy: On a standard head model, the proposed system had a mean Hausdorff distance error of 0.5443 cm, compared to 0.1868 cm for the professional FastSCAN. While the professional tool is more precise, the authors' error was less than 2% relative to the model's diagonal, which is more than sufficient for most consumer and ergonomic applications.
- Qualitative Success: The system successfully captured fine facial details (nose, eyes, mouth) and outperformed the Cyberware scanner in back-view comparisons due to better hair reconstruction.
- Efficiency: The reconstruction runs at 15 frames per second (near real-time). The physical 3D printing took between 7 and 12 hours depending on the model size.
Limitations and Future Work
- Stability: The authors noted that the Kinect v2 is sensitive to movement; they suggest adding a physical stabilizer (gimbal) to the robot in future iterations.
- Cloud Offloading: Currently, the system requires a tablet (Surface) mounted on the robot. The authors propose moving the heavy computation to the cloud so that the robot only needs to stream raw data.
- The "Last Mile" of Automation: While the digital pipeline is automatic, a human still needs to physically clear the 3D printer bed and press "start." The authors hope for future printer APIs that allow for a truly remote, hands-off interface.
Summary updated