Using Unique id and class to count objects

[quirrelHK]

Search before asking

• I have searched the Supervision issues and found no similar feature requests.

Question

In all of the implementations I saw, we are using a line or a polygon to keep track of the count of objects. But let say in case of traffic surveillance for a intersection where we have to keep count of all the vehicles(possibly a counter for each category of vehicle).
Then how can we count all the vehicles belonging to different categories?

Additional

If we have unique id for each object detected and we are also detecting the class it belongs to, then can we not just use the unique id to keep count of each object along with the class it belongs to?

[github-actions[bot]]

Hello there, thank you for opening an Issue ! 🙏🏻 The team was notified and they will get back to you asap.

[SkalskiP]

Let me move that issue into our Q&A section.

[SkalskiP]

Hi @quirrelHK 👋🏻! The short answer is yes. You can use tracker_id to count how many unique objects were seen on video. The problem is however that depending on your use case you may experience tracker_id instability.

This is especially happening when you have crowded scenes. Or scenes where there is some form of occlusion when an object disappears behind some other object and reemerges on the other side with a new id. In those cases, if you simply use the count of unique ids you will be way off.

Counting objects crossing the line is much easier from that perspective because your tracker_ids must be consistent only locally when objects cross the line. Not globally on the whole frame.

[quirrelHK]

Thank you for your reply. I couldn't connect earlier because I was stuck in a situation so didn't saw mails and naively created new issues so sorry about that.
So I actually implemented this and it was just like you said my count was way off than the observed one. Can we do something so that id switching is less if not none. Because in some cases I found that it was not possible to create this sort of line for some unstructured environments.
Thanks

[SkalskiP]

In my experience especially if we talk about long videos or streams - for example someone expect you to come up with total count from whole day - it is very hard to have result within 10% of real value. And it is almost always better to go for line counter, because like I said it only forces you to have consistent ids locally.

As for your question - can we do better? You can always do better but you need to manage your expectations. Do you need to process in real time?

[quirrelHK]

I my case for only a 15 min long video, the algorithm barely managed to keep the results within 10% of the real value. So I understand now that it is not a great idea to use id's for keeping count. Furthermore, in my use case I have thousands for videos of different locations from different angles so it would become very difficult to draw lines for each one of those.
And also in some videos the camera is not stable so I cannot use lines to count.

I will try to do better going ahead with this, but with what you told me I don't think I would get better results.

Right now I am not processing in real time. But it would be better if I can process in real time.

[SkalskiP]

Don't get me wrong. There are things you can do to improve your counter results, but it all depends on the scene you are trying to solve. Unless you are using single-stage trackers like FairMOT (where your detector and tracker form a single neural network), you probably have separate detection components and separate tracking components. In a setup like that, tracking success is pretty much dependent on the detection result. Low-hanging fruit is to work on the detector:

• Eliminate double detections
• Eliminate false negatives

[AT9991]

Hi Piotr, In your candies tracking and counting video, suppose you have two different classes of candies, say blue candy and red candy. How can I count the two classes separately. I have trained the YOLOv8 model to recognize these two classes. Can a single line detect the two classes separately? Can you please provide a workaround? Thank you!

[quirrelHK]

I am using separate components. In detection part I am getting a few double detections and false negatives as you just said. So, there is some work to be done for improving the detection part. In addition to this there are some id switching in tracking part.

So like you said first I will try to solve the issues with the detection part of the things and then also work on the tracking part.
Thanks @SkalskiP

[SkalskiP]

@AT9991, the quick workaround is to have two lines - one for each candy type. Then you filter detections like this:

detections = ...
detections_0 = detections[detections.class_id == 0]
detections_1 = detections[detections.class_id == 1]

And use detections_0 to trigger one line and detections_1 to trigger the second line.