Subscriber Discussion

Tracking A Person Through The System?

So I have told my integrator that I want the ability to highlight a person in a field of few and be able to track that person through the system. Example: If I highlight a person on camera 43 and they are moving from left to right in the field of view I want the system to know that the next camera is 40 and bring that camera up. Does anyone know if this is possible?

Edward, there are a few VMSes that let you associate cameras together like that. It's not magic (the VMS can't just predict it). You need to manually configure these associations (i.e., in the setup for camera 43, mark the right side as linked to camera 40). Then on the live view, you could click on the right side of camera 43 and then it would automatically switch you to camera 40, without the operator knowing the name or ID of that camera.

We covered this in the VMS mapping shootout under the term 'suspect tracking.' Examples include Genetec's Visual Tracking and Milestone's Camera Navigator.

Those who are aware of others, please share.

Controlware used to have something called Image Content Navigation Image Content Navigation, similar to Genetec Visual Tracking feature. ICN

It's sounds like you are asking for the system to automatically track a person across cameras and when they disappear from the FOV on one camera and pop up on another, that camera view is automatically called up.

This is not feasible in the current marketplace. For the most part the analytics technology isn't there yet, and in any semi-crowded area your "view" of the person (size, perspective, white balance/colors, resolution/detail) is going to change dramatically camera to camera.

Like any such application, you can contrive demo scenarios or selected cases where it can be made to appear workable, but it will not be reliable in a variety of everyday settings.

I fully confirm, perfect answer! Multicamera tracking it's still a research challenge, when you don't want to go in the "lab experiment" (or "commercial trick" if you prefere...;)..) but you really mean to do it in real life and generic conditions. As written correctly, color, shape, contours and any feature used by the trackers are drammatically changing camera by camera and to find a reliable way to do it it's still officially considered an open challenge in the scientific state of the art of video analysis.

You can try to approach something in case of "heavy" hypothesis a priori, that it means for example 1-2 persons in the system:

1) By cameras-map calibration you can approach an inertial strategy, tracking the person on the map and assuming that this guy exiting here and entering there, since his trajectory in the map is almost continuous, he should be the same one. Not available maybe in the market, but possible to develop it and reliable. But if I have more than 1 person interacting in that area, of course it will be much harder to decide correctly.

2) Simply connecting the analytics alerts in the VMS, automatically and in real time showing the cameras giving the "intrusion" alerts and correlated live view. This is a trick of course: the effect is to have the view of the same guy going camera by camera, but of course it's not a real multicamera tracking working. And of course this is working only with 1 person in the whole system.

Which kind of environment is in this case? For sure, for example, if the request is for tracking a customer in a commercial center or a criminal in a bank building........tell to him to forget it for many years and to beware to whom is proposing it to him!...;)




Not to get in the way of analytic vendors kicking down analytics, but as I mention above, configuring VMSes to associate cameras is the way to go here. Save for the initial time to make these links, it should work fairly well and does not depend on a computer vision miracle.

I'm not familiar with how the camera associations are configured.

Would associating cameras left and right/forward and back automatically pop-up the view from the associated cameras once a target enters the first camera's FOV, both left and right?

I'm hoping the tracking software used by the vendors that offer it is smart enough to pop-up only the camera to the right when a target is moving left to right, otherwise I can definitely see how that might be a inderance with all the pop-ups that might be generated as the target moves in an out of the FOV of each camera, although that might only be a problem when the target enters the first camera's FOV.

Have a look at Snap Surveillance. This software has algorithms that associate cameras close by automatically. Can even assess a large system and show what's not working and not associated using business intelligence...

Syd, you may not get the US questions right, but you are sure on top of Australian companies! :)

That's an interesting company and they are aiming to do that automatically using analytics. Thanks for sharing.

Here's their conceptual video, which does a good job of showing the hypothetical benefits (though it is a wildly optimistic scenario):

And here's their demo video - there is no audio but it shows the actual UI and the workflow of the software:

My main questions are: how much setup does this take? how accurate is the tracking / patterns? how much does it cost? how do you make it work with existing VMSes?

I've sent an email to Snap to ask for more details. We will do a post on them.

That's very interesting.

The only drawback I can see, if it really does what is says it does that well, is that when the operator is tracking one target, he might lose track of everything else that's happening at that time, seeing that the camera flow fills up the entire screen, which seems like a big risk depending on what the environment and the situation are.

Obviously, that would also depend on the number of operators available at any given time.