Other vehicles incorrectly detected as cyclists/two-wheelers?

Some interesting things to look at.

Poles and signs are usually OK, but a tree/shrub might be an issue to keep an eye on.

I see that there is an increase of two-wheelers in one direction, but during evening peak hours, so is that not possible? Please bear in mind that the category on the site is currently “two wheelers” (for simplicity and historical reasons), and there may also be an increase in motorbikes of various sorts (in my case, there’s an increase in local food delivery motorbikes in the evening).

The accuracy of the pedestrian and two-wheeler category on the V1 was lower than the S2, and so the numbers are harder to compare to each other - but each one should give you good indications of local trends. The speed issue may be similar, now that vehicles are maybe differently classified (we have more categories now). If you’ve only just upgraded, please give it a reasonable time to get a good baseline and we can look at this again.

We will probably cover this in more detail if you are able to come to tomorrow evening’s Telraam Talks Live with @kris.vanherle and myself.

I’ve been using my on-the-phone time to watch the Telraam count today. It’s not been for long, maybe 5-10 minutes, on two separate occasions, but each time, I see it counting two-wheelers that aren’t there. In one instance, it counted three two-wheelers when only one passed, and in another instance it counted one two-wheeler when there were none.

When I saw it, I tried to film with my phone in the hope that it would happen again immediately, but it didn’t.

It’s not windy today. The trees are almost completely still. There is a little movement, but nothing I would expect could be misinterpreted as traffic.

That is interesting to hear.

One thing we could do, if you want to help us investigate, could be to record a 15 minute video from as close to the same location as possible with a camera like a gopro (wide angle).

We could then run this through our training process and investigate if there are anomalous readings coming up and maybe train the algorithm further.

Do you own a GoPro by any chance?

I can easily do that.

After I posted the message today, I set up my 360 camera to try and record both the street and the Telraam screen at the same time, to see if I could capture more phantom cyclists. I haven’t looked through that yet, but I’ll let you know.

1 Like

The screen of the Telraam wasn’t readable (too dark), and because I put the camera slightly behind the window sill, the near sidewalk also wasn’t visible. Will set up two GoPros tomorrow to record the street and the Telraam screen.

Also want to highlight the car count, ref my previous comment. So far today, the standing is:

Telraam → 11566
Offical counter → 11733

There’s no strong indication it’s misclassifying cars as bikes.

1 Like

I have one hour video of the street and the screen of the Telraam. I’ll look through it and see if I can identify miscounts and crop the video down to those segments.

Do you want the raw files regardless of any identified miscounts?

Yesterday’s car counts are now final, and there is some difference.

Telraam → 12949
Offical counter → 15388

There is a potential for cars to have been misclassified as cyclists.

It is more likely that our “cars” is more restricted than the counter, as we separate out other types of vehicles and some may have been added to the “heavy vehicles” category?

In terms of the video, it might be best to send us the file to look at and analyse rather than sections, unless there’s a 15 minute section that has particular issues?

I’ll PM you my email and maybe we can use a large file transfer service to get that from you :slight_smile:

That’s true. The official counter counts cars as less than 5,6 m length. Your cut-off may be different.

ah this is superuseful feedback, thanks!

I’m a bit puzzled on the bike over-counting @balchen. That shouldn’t happen. Undercount is possible because objects aren’t always recognized as bikes or there’s 2 close to eachother, missing one.

All I can think of is that the tracker “looses track” of the identified bike and thus counts a single bike as 2 separate bikes. We didn’t see this when we were testing though, so something is weird. Is there a tree/shrub in the field of view, potentially “cutting” the tracking? Like Rob said, poles aren’t an issue, but shrubs can be. Cars are bigger, so risk is smaller there.

A video would be great, becasue we can test “what Telraam sees” in our offline pipeline, for debugging purposes.

Happy to discuss further this evening in Telraam Talks live! :slight_smile:

1 Like

There are small trees, so overcounting a single cyclist because it disappears and reappears is certainly possible. The overcounting looks like it is happening on the far side of the street (based on my observation that one directon always has more cyclists), where the trees would potentially be in the way. There are no trees between the camera and the near side of the street.

But I also see it counting cyclists when there are no cyclists.

I’ve uploaded an hour long GoPro video of the street that is captured 10 cm from the Telraam. @Rob_Telraam has the URL for it. That should show you everything. I will look through the videos and see if I can see it counting phantom cyclists.

3 Likes

Another “thing” has happened to the traffic count. Maybe this belongs in a separate thread. From (and including) June 2nd, the counted traffic number has halved. When I reference the offical counter down the street, nothing has changed to that count.

Counting still happens in both directions, and for all hours of the day. The trends looks roughly the same, only half the magnitude.

Is there a logical explanation for this?

The weather has been very sunny the past week, but the window is only in the sun after mid-day (west-facing), and the hourly trend doesn’t show a drop after the sun hits the window compared to the hourly trend before June 2nd.

Actually, there is a difference in the hourly trends… Could it be sunshine after all?

Before:

After:

Hi @balchen, we ran a “debug-video” for your device. Basically what we do, is to run a short video trough the same detection model of your device (but offline and only for the short video you provided). We overlay the detection-script and can visually see what the device is counting. This is addressing both your question on bike overcounts and your latest finding on missed detections with the sun now. A few findings:

  1. First and foremost, this site is quite a challegning location for Telraam. The street is very wide and there’s some obstacles in the field of view. I would have expected missed detections on the far side, but that’s not the case; it’s even counting pedestrians & bikes on the far side. so far so good.
  2. The cause of the overcount seems to be detection of static objects (bikes/pedestrians) which are “picked up” by the tracking algortihm and then counted as moving, adding to the tally. Those objects are near the bus-stop and are on the edge of the region of interest. Our hypothesis now is that the recognizable static objects, in combination with their location on the edge of the RoI is causing these extra counts. I can’t really see what can be done to avoid it. You may want to consider a different RoI, trying to get the bus-stop out of the RoI. However, no guarantees it will be better.
  3. undercounts due to direct sunlight, yes, absolutely. If the orientation of your window is S/SW and you have episodes of direct sunlight in your Telraam, that will cause problems. I would expect mostly (only?) the far side is affected. It may be worse in spring/autumn, depending on lattitude and window orientation. You could try shielding the device, but should be from the outside and I understand thisis not practical. Actually, if you can shoot another video at a time & conditions it’s the worst, we can run a 2nd debug video.

Overall conclusion from what we can see, quite a challenging location and a bit a of “bad luck” (bus stop on the edge, direct sunlight due to orientation). I’d consider trying with different ROI’s. The nice thing is you can directly compare with a high-end traffic counter nearby and see what setting yields the best results (which makes this case super useful for us as well!). Secondly, not sure if it’s an option, but try installation on a different window (a bit lower and -if possible- with a cleaner field of view).

Please keep feedback coming, this is really usefule for us, so we know what to focus on in the next development iteration!

3 Likes

Thank you for the feedback, @kris.vanherle.

To address your last suggestion first; I did try it one floor down, and it was noticably worse. It couldn’t count the far side reliably at all. The obstacles – I assume that’s mainly the trees – line the central busway, so there’s no better horizontal place.

The bus stop covers the entire ROI. The shelter is at the left edge, but the platform runs the whole length of the image. It’s logical that there are more people standing near the shelter. If I could freely select a ROI, I would probably select it differently, but there seems to be no way to do that. As it stands, I can’t exclude the bus shelter while including the far side of the street.

I can’t imagine many cyclists standing still at the bus shelter, though? Also, when I look through my demo video, I see it counting when there are no cyclists and no one standing at the shelter. Could there be other explanations as well?

The sunlight issue has subsided. I don’t know why. The window still gets direct sunlight in the afternoon, but whatever caused the undercount has decreased significantly. Maybe a specific angle of direct sunlight is dramatically worse. Or maybe something else entirely.