3D human body scanner

Problem

3D scanners have moved beyond labs and factories, but a good full-body scan still usually needs expensive kit. The brief for our software engineering project in the Computer Vision master’s was to build a home-made scanner instead: off-the-shelf Kinect v2, a turntable we built ourselves, and software that turns raw depth frames into an accurate, watertight 3D model of a person in under 90 seconds of acquisition.

What I did

I was the project manager for a team of ten, supervised by Dr Yohan Fougerolle and Cansen Jiang. We split into four sub-teams: scanning and registration, meshing, GUI, and hardware. We ran it with weekly reports, a Gantt chart and FreedCamp, and moved effort between teams as the bottlenecks shifted.

I also led the meshing team. I interfaced the Kinect v2 with the Point Cloud Library through a dedicated grabber, since PCL had no native support, and implemented and compared three surface reconstruction methods: greedy projection triangulation, grid projection and Poisson reconstruction.

The full pipeline grabs point clouds from the stationary Kinect as the person turns on the table, cleans them with pass-through, statistical outlier and voxel-grid filters, aligns them with ICP (including two-stage ICP and ICP with normals), smooths them, and meshes the result, all driven from a Qt GUI that also controls the turntable over serial.

Result

The software produces watertight triangle meshes from the scans, and the team delivered on schedule, with the turntable and development costing under €100. Comparing meshing methods, greedy triangulation gave the best mix of accuracy, speed and memory for small and medium objects and grid projection suited large ones, while Poisson took over 11 minutes and left too much surface to clean up.