Quividi’s software support any USB camera that allow the UVC norm (Standard USB Video stream norm) and any IP camera supporting MJPEG over HTTP, H264 over RTSP or (starting with 5.7.0) RTP over multicast UDP.
Position & Tilt
The best way to position a camera is to make sure it is placed as close as possible to a person’s eye-level.
Generally, anything between 1.60 meters and 2.00 meters would be great. (Between 5″3 and 6″6).
Quividi’s software has been trained to detect faces that are looking straight to the camera.
We recommend that you do not tilt your camera more than 15° up or down – this way you ensure that detections won’t be skewed.
Camera field of view & resolution
The Quividi solution requires a minimum of 24 by 24 pixels in a video stream to detect a face. Once a face is 24×24 or higher, the solution will detect it, track it and classify it by demographics. The maximum detection distance is the limit beyond which a face is smaller than 24×24 pixels.
The Quividi solution’s detection zone is defined by the camera resolution (ie how many pixels it is ingesting) and the camera’s horizontal field of view (ie the angle that it covers).
The higher the resolution, the further you’ll be detecting faces. However, the higher the resolution, the more processing power you’ll be need.
The broader the field of view, the more “on the sides” you’ll be detecting faces. However, the broader the field of view, the further the people will appear to be, and the shorter you detection distance will be.
Think of it as a slice of a tart: for a same quantity of tart (i.e. of pixels processed), you could either have a thin but long slice, or a fat but short slice.
A specific section of this tutorial is dedicated to estimating the detection distance and minimum CPUmarks (ie processing power) needed, depending on your settings.
Position vs. traffic flow
The more perpendicular to the global human traffic, the more watchers VidiReports is going to detect. Indeed, when they enter the detection zone, people walking towards the screen will hardly miss it and their face will be detected and the conversion ratio of watchers divided by passers-by (also named Opportunities to see) will generally be good.
Conversely, if the screen is parallel to the main flow (such as when a screen is against a wall in a corridor) and the camera is aiming right in front of the screen, then people will generally not have their face turned towards the screen as they enter the detection zone. In such cases, conversion ratio will generally be low.
So, to get an optimal number of watchers, you’ll want to set your camera (and possibly your screen orientation) in such a way that it detects faces when they enter the detection zone.
Note that Opportunities to See remain rather unaffected by the orientation of the camera.
Quividi’s software is only as good as the image it receives from the camera !
When the camera has to be placed behind a window that is subject to reflection, the best way to avoid the effect of reflection (glare) on the window is to stick the camera on the window.
Unfortunately there are no magic tricks, if your eye can see the reflection so will the camera. The only variables that one can play with are:
- distance camera-glass: the closer the camera, the fewer the reflections. Ideally, glue the camera to the glass
- indoor light. If the exterior is darker than the interior, the reflection will be more apparent. Avoid using a very bright screen behind the glass, reduce the brightness as much as possible.
- clean glass: dust or other films may increase the reflection
- polarized filter (like one of those): it minimizes reflections but does not eliminate them. You also need to rotate the filter to find the best result
Wide Dynamic Range
Please make sure the dynamic range of the scene is not causing any issue : if an area is completely overexposed, or underexposed, the software will not be able to detect any faces.
Recent cameras usually have built-in software capabilities to minimize the effect of strong lighting / dark shadows, but you should always double-check that it works in your installation.