I installed my Telraam v2 today and have watched it as it counts. I don’t know if the numbers on the screen are real-time or not, so that may have some bearing on this, but maybe not.
As I looked at the screen, I saw the count for cyclists/two-wheelers increase by 10 even though no two-wheelers were passing. This happened over several minutes. At first I reasoned maybe the count was lagging behind slightly and that I was seeing a count for cyclists already passed, but was not expecting such a lag to be several minutes long.
At the moment, when two hours have passed, the two-wheeler count is also far higher than I believe from my observations.
What am I seeing here? Is it a classification issue or something else?
The data display should not lag that much, it is only a matter of seconds. It could be that anomalous movement is creating false readings. Can you tell us more about your set up and your view? Have you selected the most appropriate Region of Interest already?
ps. welcome to the community, and thanks for sharing
My view is from the third floor (second floor in British terminology) with a fairly unobstructed view of the street. There are some small trees and light poles, almost everything on the street is clearly visible.
The street itself is not easily described with the street layout tool, since it has a central, two-way busway flanked by one-way roads with cycle lanes. But I think that shouldn’t affect the count.
I set the ROI when I installed the device. It was a bit tricky to pick the right frame – if I could have cropped the image myself, I would have chosen a different crop than the ones presented to me. This is probably because the street is so wide. The ROI now includes an access road for adjacent businesses opposite the street from the camera. That means my numbers might include the access road, but it does not have a significant amount of traffic compared to the main street.
I’ve since read the details on how classification is done (axis and fullness), which in the description seems to be based on a horizontal view of the objects to be classified. Being on the third floor, my view is probably around 60 degrees for the nearest sidewalk and 30 degrees for the farthest sidewalk.
Apart from that, I can only say that Norwegian cyclists generally are not riding as up-right as the example image of a cyclist, but that should probably affect the count the other way around
thank you so much for this information - I will take a look with the team and see whether we can give further insight.
However, just to quickly answer the question about calibration, the S2 should not need to calibrate at all. The calibration period warning relates to the V1 as it worked differently.
However, as with all data, you need to collect a longer period of data to be able to make any decisions on the reports as there could be many reasons for variations in a particular hour / day, but these should smooth out over time. The main reason to check in the short term is mainly to see that it is positioned correctly, so we can look into this and see if we can provide any further suggestions.
It does seem to have “landed” on more realistic modal shares for the last 24 hours, hence my comment about calibration. You can still see in the graphs that the cycle/two-wheel modal share was way higher the first hours it was in operation, and because I was observing, I can tell you it was not because of an actual influx of cyclists.
Also, so I’m not just complaining: there is an official traffic counter just south of my location, with just one access point in between, so the numbers should be directly comparable. The official counter shows 9932 cars today. The Telraam shows 10557. That’s very good
The official counter also counts cycles. That count is 468, vs. 1357 on the Telraam. But the Telraam will count people cycling on the sidewalk, and there are several access points for cyclists between the two counters. And this is part of the point of using the Telraam – to see what the officials are (probably) missing.
Ah, now that is very interesting. Thank you for this information and background. We’d certainly love to know more about how the data comparison goes over time.
Indeed, we should, in theory capture more cyclists in particular since the cycle-specific counters have limitations (as you usually need to cycle over specific bits of the street). Let’s hope it settles down in that case.
I’ll keep an eye on it and see how it tracks over time. If the initial glitch in counting has stopped, it will certainly not be a problem. I just want to prevent it from happening again, as far as possible.
Do you have any input on the placement of the device, third floor vs. second floor?
There are some small trees visible in frame. Probably 4 m tall, 1 m wide crown. If it’s very windy, could their back and forth motion be counted as cyclists due to their shape? It was very windy for a time, and we do get a lot of wind here from time to time.
I think I see from the graphs that the phantom cyclists were counted in the same direction (B->A). I don’t know if you agree with that. If you do agree with that, does that tell us something about the cause of the excess count?
It is not obvious what the issue might be from the data alone - in fact one of the upgrades we are working on will be the ability of hosts like you to annotate data to point out issues and give explanations for anomalies - since you know the road best, more than us!
Let’s keep an eye on this. The road is quite wide and the placement is fine, but objects like pedestrians and bikes on the far side will be quite small for our camera, but if anything this should result in undercounts.
If it persists, then we can chat further as one option would be to send us a video that we can analyse and see if we can discover any issues, and maybe even re-train our algorithm to take this into account - but we’re not there yet!
Since the majority of cyclists counted are B->A, they would naturally be on the opposite side, so there’s no reason to think it’s undercounting the opposite side. My main concern in that respect is the near sidewalk. I’ll see if I can either tilt the camera slightly or move it down one floor to get a better view of it. Right now the camera only shows half of the sidewalk.
I moved the camera down one floor this morning at 7. The results from today show me clearly that it’s no longer able to count the opposite side of the street. I did re-set the ROI immediately after the move, but to the same value as was already configured, since it still covered the desired area.
After seeing the issue, I’ve re-set the ROI to the next wider view (from bottom 7 to bottom 6) to see if that helps. If it doesn’t help, I might need to move it back up one floor.
Thank you for this thread my newly set up S2 is also representing a high number of two wheels (and whilst I have set it up on the outskirts of Cambridge UK) and would love to see that number of cycles over the number of vehicles, there is something else coming into play. Device Telraam | Addenbrookes Road
I am also second floor town house position of the device due to access to a power source. Have set the ROI but do have a road sign, street light and tree in the selected area. The shrubs do obscure the main cycle/pedestrian path and the road is a challenge to define. One lane east to west (A to B) which just prior to the ROI there is a short section of dual lane where drivers are meant to merge but due to the wide single lane do have penchant for racing each other or failing to yield as must get in front and one west to east (B to A) has a number of drivers using the central hatching as an additional lane to race the lights at the intersection further along. Thank you for any assistance.
The good thing about my setup is I can verify the car count based on the official car counter further down the street. Save for the time when I had the Telraam one floor down, it has been counting cars almost the same as the magnetic counter.