Trim3D: Development of an Automated Hair Trimmer

This bachelor’s thesis presents the development of Trim3D, a prototype for an automated hair-trimming
system. Starting from the observation that hairdressing remains mostly manual, the project aims to make the
first step into automating the haircutting process and addresses core challenges, such as accurate 3D head
measurement, intuitive haircut definition, and precise real-time control of a hair trimmer with an extendable
comb.
The developed system integrates a three-camera network for synchronous multi-view capture (requiring
intrinsic and extrinsic camera calibration with ChArUco boards, and triangulation to determine 3D positions
from the different camera angles), ArUco marker-based pose estimation for head measurement and trimmer
pose estimation, and MediaPipe facial landmark detection for head pose estimation.
After rigorous systematic evaluation and testing (often using Blender to ensure a controlled environment,
availability of ground-truth, and reproducibility), the biggest error sources were identified, and the software and
hardware were reworked or alternative approaches were found, evaluated and implemented. Through targeted
improvements, such as multi-threaded image grabbing, triangulation reprojection error thresholds, ArUco
dictionary changes, ArUco marker pose ambiguity correction by global marker rotation-alignment across
cameras, thin-plate-spline head morphing with outlier handling, image exposure correction for more robust
facial landmark detection, stability-enhancing mechanical redesigns, and 10 Gbps networking with jumbo
frames, the system’s accuracy and robustness were greatly enhanced. Similarly, the performance increased,
doubling the system’s frame rate from around 10 to 20 FPS. End-to-end tests of the Trim3D system in a
simulated environment verify the detected trimmer position to be accurate with errors below 5 mm and
predicted hair length versus ground truth hair length with errors below 0.5 mm at different positions.
Unifying all underlying components together, a web interface enhances usability and guides users seamlessly
through extrinsic calibration, head measurement and morphing, haircut definition and real-time trim execution.
The final prototype demonstrates feasibility, accuracy, and user-friendliness, laying the groundwork for future
visions such as smartphone-based head scans, sensor fusion, and an actively tilting trimmer comb.