How many cameras and where
Calibration does not depend on the number of cameras, reconstruction does; staggered heights, minimum simultaneous views, a subject leaving the scene.
Transcript
1. The camera ring
The cameras are on the walls of a 6 by 6 metre room, evenly spaced along the perimeter and aimed at the centre of the 3.5 by 3.5 by 2 metre measurement volume.
To frame the whole volume from so close a 120 degree lens is needed. We try 3 to 12 cameras.
The strip below is the “unrolled” perimeter: it shows the height of each camera.
2. The heights
Three layouts: all at the same height, 2.2 metres; alternating on two rows, 2.4 and 1 metre; rotating on three rows.
The hypothesis is that staggered cameras see the legs better and cross their rays at wider angles.
3. Calibration and number of cameras
First surprise: in the part of the volume crossed by the T the calibration error is between 0.3 and 0.45 mm with 3 cameras as with 12.
Each camera is calibrated by the same observations of the T: adding cameras does not improve the others. Camera errors stay below 2 mm in position and 0.04 degrees in rotation in every case.
Where large errors appear, it is because the T did not reach the corners of some camera’s image, and there the lens distortion is extrapolated.
4. How many cameras see each marker
To reconstruct an athlete, instead, the number of cameras matters. Thirty markers on the body of a walking person: how many are seen by at least 2 and by at least 3 cameras?
With 4 cameras only 25% of the markers are seen by 3 cameras; with 8, 91%; with 10, 96%; with 12, 97%.
Reconstruction recall follows the same curve: 48% with 4 cameras, 97% with 8, 98.7% with 10. The gain converges between 8 and 10 cameras.
5. Staggered cameras
Marker reconstruction error for the four height layouts.
With 8 cameras: high row 1.33 mm, two rows 1.16, three rows 1.15. With 10 cameras three rows reach 1.01 mm. The gain, 10–20%, comes from the low cameras.
The hypothesis is confirmed: cameras staggered in height improve the reconstruction of legs and feet.
6. Simultaneous views
How many cameras must see the T at the same instant? We keep at most k cameras per frame and delete the others.
With a single camera per frame the calibration fails: the cameras are linked only by the terna, seen from far away.
With 2 it works, with 3 it is already close to the case with all of them. Minimum 2, recommended 3: the T must pass where cameras overlap.
7. The subject leaves the scene
Whoever moves the T sometimes leaves the volume. Blue path: always inside. Yellow path: up to half a metre from the walls.
The histogram counts, frame by frame, how many cameras see the T. Leaving the volume increases the frames seen by few cameras.
But the calibration does not get worse: frames with fewer than 2 views carry no information and do no harm. It is enough that in the rest of the session the T crosses the centre.
8. Recommendations
In short: 3 cameras are enough to calibrate; to reconstruct an athlete at least 6 are needed, better 8–10; beyond 10 the gain is below 1%.
Cameras staggered in height, at least 2 simultaneous views of the T, and cover the corners of every image too. In the next video: the algorithm.