Accuracy counting pedestrians and cyclists

Hi – we have recently installed a Telraam and it seems to be systematically undercounting cyclists and pedestrians; are there any tips we should take into account to optimise its performance; e.g. is it a problem if it is two floors above ground level?

it will depend on the view from that location. You have to remember a couple of things:

  1. it is a camera, so it needs to see the objects, but it is low resolution so they need to be reasonably clear. If they are obscured or too distant, the device may struggle

  2. it is an AI, and it is trained on object largely seen from the side. A pedestrian looks rather different from the side and from above, so the angle will have some effect

  3. it is a device that has been designed to be cost-effective and sufficient to get useful data over time. It cannot guarantee 100% accuracy on any category, but particularly for small objects like pedestrians or bikes, especially when they may be obscured by walking / cycling in groups.

There are many reasons counts will be under an absolute number, but since they are systematic and continuous, you will get a very good picture of trends across time and over periods like days / months / seasons, etc.

One thing to check is whether you have set the ROI - as this can help the device to focus on the key areas and even zoom slightly if they are further away.

Lastly, if you want to send us a validation count, that will always help to decide how different the counts are and whether there is anything more we can suggest:

Take a look at our counter setup here in Provincetown. It seems to do a very good job counting pedestrians as long as there isn’t anything obstructing the view of the street (occasionally a car parks in the street and the counts drop precipitously). I’ve done a number of manual validation counts to confirm this.

1 Like

Thanks for the comments – modifying the field of view has definitely helped with the pedestrians.

1 Like