This article explores the selection of a Region on interest for the Telraam S2 device. Further in this article you can find:
An explanation on "How to select your ROI?".
Examples and guidelines for a "ROI selection".
Why an ROI on Telraam S2?
One of the big quality of life improvements of the Telraam S2 sensor compared to our first Raspberry Pi-based Telraam is the fixed, built-in, wide-angle camera that replaces the admittedly flimsy external camera unit of the earlier device. This eliminates the risk of a separate camera unit coming loose and inadvertently tilting away from the street, therefore making the need for capturing daily background images obsolete, as the field of view (FOV) cannot change anymore.
The region that the camera unit of a Telraam S2 can see extends from nearly all the way vertically down, to slightly above the horizon. As a result there is practically no part of the street where it could not detect road users, with the exception of people passing right underneath the window, but at such angles the chances are high that part of the window or the building would be in the way anyway.
In most cases, however, this wide field of view is unnecessary, as traffic occurs in a much narrower range, and Telraam does not need to monitor areas which are free of traffic, such as the sky, the buildings across the road, or a front garden in front of your house. This is why Telraam S2 has the capability to restrict the FOV to a more limited area, which we call the region of interest, or simply the ROI. While Telraam S2 can automatically select the ROI it thinks is the best, users are advised to fine-tune the ROI selection themselves (see examples at the bottom of the page), as it has many advantages compared to an automated setting:
A manual ROI setting is permanent (until changed), while an automated selection by the Telraam S2 will happen every time there is a reboot (after a power cycle or a firmware update), and each automatic setting might be slightly different, and in some cases might be sub-optimal.
A manual ROI setting will provide visual feedback on what the Telraam S2 actually sees.
An optimally set ROI will provide the highest accuracy numbers in terms of road user counts. The AI in the Telraam S2 unit uses the same (low) image resolution no matter what the ROI setting is. On a narrowly set ROI a pedestrian might be 4 pixels tall, while on the widest ROI it might be less than 2 pixels. This means that to have a better resolution of small road users, such as pedestrians on the far side of the road, choosing the narrowest ROI that still includes all road users along with the least amount of static area is the best.
How to select your ROI?
The ROI can be set via the "Camera Images" section.
The dashboard will show if no ROI is yet set. Click the button "Set Region of Interest (ROI)" to carry out this action. A wide angle image will be requested and you will be asked to select one of the zoom options shown as coloured squares.
ROI selection examples
To better illustrate good ROI-selection practices, we have created two example mock-ups using higher resolution images from a GoPro camera.
Example #1 below shows a street where the road surface is close below the camera, so we need to select one of the ROIs which are looking at the lower portions of the image (6 to 10). The optimal ROI in this case is 8, as it covers all areas of the image where traffic can be found, from the sidewalk at the very bottom, to the sidewalk on the opposite side of the road, behind the parked cars. Choosing a narrower ROI (such as 9 or 10) would mean that we cut out the traffic on the far side of the road, while choosing a wider ROI (such as 7 or 6) would only include traffic-free, stationary parts - such as the houses on the opposite side - on top of what ROI 8 already includes.
Example #2 below shows a situation where the road surface is much further away from the camera, and a large part of the widest possible ROI - practically the bottom part - is static, and will never contain traffic. The optimal ROI is number 4 here, which zooms in at the top half of the image, leaving out as much of the static foreground possible without cutting into the areas with traffic.