Camera Setup

There are a few things to consider when planning a new camera setup. The hard constraints can be summarized by a few constraints:

1. Make sure the resolution is high enough.
2. Every point in the volume needs to be visible from multiple directions.
3. Maximize the angle between cameras.
4. Keep cameras above head-height but below 60°.

Resolution

For body tracking without fingers and optimum tracking performance, we recommend that one pixel in-camera represents no more than 10mm/px in the real world. That means, at a resolution of 1440px x 1080px the maximum width of the field of view of the camera should not exceed 1440px * 10mm/px = 14.4m. If the cameras have significant lens distortion the effective resolution is reduced. So you may want to restrict yourself to a lower horizontal field of view.

For finger tracking the resolution requirement is a bit higher. We recommend around 5mm/px or 7.2m horizontal field of view at a camera resolution of 1440 x 1080 pixels.

For face tracking an even higher resolution is required. We need the face to be no smaller than 250 pixels. So you may have to restrict how far a person can be from the most frontal camera.

Camera Count

Generally, we recommend camera setups with 8+ cameras. However, this includes some redundancy for camera failure and occlusions. Tracking quality does not degrade if only 6 cameras can see a person and as long as 4 cameras can fully see the tracked person, we don't expect significant tracking quality degradation. Do expect some degradation when going below 4 cameras. With only one camera being able to see a person, tracking will become unstable and problems with occlusion (e.g. that an arm is not visible any longer) will occur frequently.

Camera Placement

Ideally, in order to maximize the amount of information a camera can contribute to the tracking, you want to maximize the angle between cameras. Taking this advice to the extreme, you'd end up with an evenly spaced hemispherical setup. However, there are additional constraints:

• Our AI system was trained with cameras looking down at angles ≤ 60°. They do generalize well beyond these 60° but if more than a few (one or two) cameras are looking down at a steep angle tracking quality, especially of the legs and arms can degrade. There are situations, where some high cameras can improve tracking though.
• Practically, it's good advice, especially in publicly accessible areas, and in multi-person scenarios to keep cameras above head-height. The likelihood of cameras getting bumped is significantly reduced and occlusions between tracked persons is diminished.