In order to reduce false alerts and improve tracking accuracy, an estimation of object size in meters is utilized (calibration). Typically, automatic calibration is sufficient for many scenes with high object traffic. However, in more sterile scenes or those with uneven, multi-level terrain, a manual calibration may be required to generate a pixel-to-meter map. Note that the system will automatically generate an insufficient calibration warning if it detects that the automatic calibration has been insufficient for longer than seven days.

The manual calibration is done using vertical markers with a defined meter height. Each marker should signify the approximate height of objects in the scene at different locations in the image. Once the manual calibration is saved, the analytics service agent will convert the manual calibration marker positions and heights to a mapping of heights across the scene. This mapping will most closely match the user-defined markers. As such, the number and position of markers defined should vary according to the scene.

Manual calibration markers are treated as the ground truth for the calibration, meaning that they should be placed at the most critical locations in the scene for accurate identification of objects. For the model to best map the scene it is also beneficial to space the markers out and not have them clustered in a small space as that will give the model a better representation of the scene. Additionally, if there are non-gradual changes in expected object height (such as walls, and multi-level areas), markers should be placed on either side of the height change (see example 2 below).

Manual calibration requires a minimum of three markers in order to work. This minimum may be sufficient for scenes with a flat, linear landscape. However, in order to accurately map scenes with nonlinear landscapes (hills, multi-level areas, significant camera warping, etc.), additional markers may be required.

Here we see a scene that is a good candidate for automatic calibration. There are enough objects moving through the scene, the road is flat, and there is minimal warping from the camera lens. If we were to look at the automatic calibration, we would expect to see 256 different markers spread throughout the scene, such as in the image below.

If we were to decide to manually calibrate, it is likely that three markers would be enough. Note that when visualizing the manual calibration, we don’t see all the automatically generated markers; instead, only the ones which were defined by the user are shown.

This example shows a multi-level situation. If we were to allow the automatic calibration to work, it would calibrate well for most areas in the scene. However, if we are interested in identifying people on the bridge at the bottom of the screen, automatic calibration would likely not be sufficient as there is very little observable traffic in that area.

For the same reason that the automatic calibration would not be sufficient to map the bridge and the various road levels accurately, the three-marker manual calibration is also insufficient.

In this image, we see a better manual calibration. Here we defined calibration markers at various positions, including the bridge. Note that markers were placed at locations where the road surface is at a different elevation to allow the model to most accurately map the area between markers.

This example shows a very sterile scene in which we want to identify and track people on multiple levels. As very little is expected to pass through the scene and we’re interested in tracking people on both sets of train tracks, manual calibration is necessary. Note that we placed markers at all heights that we want to track people, as well as the location at which we expect a significant change in elevation (the bottom of the hill). This will allow the calibration to accurately interpolate the change in heights between the markers in order to create a more accurate mapping of the incline to the higher set of tracks.