All Categories Telraam S2 Potential inaccuracies with Telraam S2

Potential inaccuracies with Telraam S2

While Telraam S2 is much less susceptible to circumstances that could cause the traffic counts to be inaccurate, there are still some factors that may lead to accuracy values that don’t reflect the expected performance outlined in the FAQ article Telraam S2 count performance validation. These situations are not expected at the majority of typical locations where Telraam-installations take place, and most of them can be avoided by evaluating - and if needed excluding - candidate locations with these limitations in mind.

  1. Extreme light situations: Telraam uses a small digital camera to provide data for the on-board AI unit, which means that if the camera cannot provide adequate images to the AI chip, then the resulting counts will be inaccurate. These situations can happen simply because the dynamic range of the small imaging sensor is small. If you see that counts from your Telraam are behaving strangely in sunny weather, there is a good chance that this is the cause. 

    • Sun in the field of view: if at some point the Sun shines directly in the field of view (FOV) of the camera, that could “blind the sensor”, by either saturating the image, or by making objects too dark to be recognised. It can be that when you install a Telraam S2 during the summer on a South facing window then this is not an issue, but during the winter as the Sun seems to travel across the sky at a lower altitude, it might become one. Note: the field of view is not the same as the region of interest (ROI), as it is basically the widest possible region of interest. The exposure control of the camera uses the whole field of view due to technical limitations, so if your ROI is set narrower, then you might not see the Sun in the ROI image, even if it is inside the field of view.

    • Extreme contrast: in some cases areas of strong sunshine (especially on bright, reflective surfaces) and strong shadow (on the street level) can create situations where everything in the sunny areas will be too bright, and everything in the shade will be too dark for the AI to be detected. This is quite rare and should not happen in typical scenarios. 

  2. Obstructed view:

    • Window blinds: as Telraam S2 hangs on the inside of your window, if you have external window blinds, then the device will not be able to work when the blinds are closed. 

    • Weather: similarly, if the windows are covered with raindrops or a layer of dirt, the view of Telraam S2 will be partially or fully blocked and the device will not be able to count correctly. 

    • Street furniture and vegetation: If trees, light poles, railings, bridges, or similar structures block part of the ROI, the AI might miss some objects or it might not be able to track them to a sufficient extent. While a single pole or narrower structures are not an issue for the AI, a row of trees, thick horizontal railings, are. A good rule of thumb is: if the structure is significant compared to the size of the street users, it will likely cause issues. 

    • Unseen traffic lanes: If there is, e.g., a bike lane behind a parked row of cars, then cyclists passing there will likely not be counted. Even if you can see people’s heads poking out from behind the obstruction, that is simply not sufficient for the AI for detection, mostly because 1) a head could belong to a cyclist or a pedestrian, and 2) a head in a low resolution image that is used by the AI is less than a pixel in size at typical distances.

  3. Non-optimal ROI setting: see our FAQ article on Region of interest (ROI) and ROI selection with Telraam S2. An overly wide ROI might lead to small objects not being detected - especially on the far side of the road -, while an overly narrow ROI might exclude part of the traffic (or cut off the top part of taller vehicles), which will lead to incorrect counts and potential misclassification.

  4. Atypical traffic:

    • Unexpected classes: Telraam S2 can identify objects that fall into the 10 categories that were defined during the training of the AI. Telraam might not be able to detect or count objects that do not fall into these categories. 

    • Crowding: very dense crowds of pedestrians and cyclists can be a difficult task for Telraam S2. While cyclists cycling side by side at normal distances won’t cause a problem as long as they don’t overlap too much, counting people in a mass-participation event, such as a protest, or a running race is too big of a task. In these cases some undercount is expected.

We expect misclassification to be much lower with Telraam S2 compared to Telraam V1, although with classes that are very similar to each other, such as larger vans and smaller trucks, or light motorcycles and normal bicycles, there is always a chance. But cyclists will never be classified as cars, and cars will never be classified as heavy vehicles anymore, while in some scenarios this was a possibility with Telraam V1.

To read about potential inaccuracies of the speed measurements of Telraam S2, consult our FAQ article Speed measurements with Telraam S2.

Was this article helpful?

Thanks for your feedback!