Motion capturing

Active: 2019

Capturing sword fighting

Trying to capture martial arts (kenjutsu) practice sequences with cheap video cameras (phones).

Wearing black clothes with colorful dots put on. Capturing from 2 different angles.

Sadly, finding the dots is only half automatic. Some image recognition helps, but often it still involves clicking on pictures.

Camera positions and orientations can be guessed from the corners of a known object.

(It also works ok, to find this information directly in the reconstruction step, but that is not very stable)

With the camera positions known, finding the 3d positions from the 2d dots can be a convex optimization problem, i.e. it is fairly simple and stable.

Afterwards, some simple filtering is used to smoothen out jittering.

Since dots are too often occluded, a setup with more cameras would make things easier.
The cameras' pictures are very distorted. Syncing with the known object in one location gives bad reconstruction 1 meter away.
Needs more anatomical information to guess bone orientation around the third axis (esp. feet and sword).
So far, the dots are simply used as the joint centers, even though they are attached outside the body (this causes the sword to be far from the hand).