Are you speaking of a hardware or software solution John. Meaning, having multiple static camera's in a store where the VMS can actually reconise an individual trough the multiple video's, or having PTZ's who are directed towards the individual you're tracking ?
I do have somewhat experiance with the latter, only this one had but 1 PTZ. It consists of 2 camera's, and uses the system you describe (one static, one PTZ). The biggest downside with these systems is that you're limited to what the static camera can see.
A collegue once tried out the auto-tracking function of one of Bosch it's dome's. Now this was a few years ago, but it wasn't exactly giving off the results one would expect.
I did see an intresting system though for a large sized area. CBC's Ganz Radar. I was quite impressed with the results of the test I saw.
Rogier, a number of companies offer a master/slave fixed / PTZ combination. That is really just autotracking++, an incrementally better way to autotrack with a single PTZ.
What I am asking is tracking across non overlapping cameras. Imagine following someone through a large mall or public area or down a road. I suspect this is mostly science fiction but I am curious if anyone is trying this out or claiming to do it.
It's obviously 'possible' to do as an R&D project or demo. The questions are: How accurate is it? How much does accuracy decline as the number of people in a scene increase? How much does environment and lightining impact it? How much 'intelligent' / analytic processing do you need on the backend to make this work? What type of costs does that add?
Reading from the paper it's quite accurate. They only had trouble with people taking way longer in the non-seen areas as they inticipated.
However, the paper is based on streets where as people move from point A to B using their shortest route, which is logical for a public street. I doubt this would work in a store however, as customers can move in irradic pattern (women in clothes stores, need I say more). This causes for potential long periods in the non-seen areas, thus screwing up the analytics.
It's indeed a shame they haven't listed their used hardware. I wonder what camera's they used, and specificly, which server hardware. It could give us a sense of total cost.
Well, from the images that they show, the scenes are fairly ideal / simplisitic, e.g.
Btw, the paper is 10 years old. Also, even if the hardware specs were current, that's just for an ideal environment, certainly it would take far more processing power to analzye over larger areas with greater environmental variances and more simultaneously subjects to track.
John, We had systems doing this in the late 90's early 2000's using Pelco and Panasonic equipment along with a standalone software package. You would tag a suspect and it could follow them throughout the mall. (i got in trouble with this one while we were testing it...forgot about the monitor in the mall office..lol) It used a Series of Presets that when the subject was reaching a certain preset it could then trigger the next preset of the camera or one on another camera...thus following the subject through the mall. Not sure how this would work in a larger open space but worked ok in the mall when it wasnt to crowded. (sorry but I cannot remember the software package name and no longer work for the servicing company) but I do remember it was a whole lot of ascii through a pansonic matrix system.
Update: Here's a research project from Carnegie Mellon that tracked individuals across multiple cameras. It's not ready from prime time, cost a lot of money and required a lot of cameras but it's interesting research into the specific issues involved.
This video shows fixed to PTZ handoff being performed by ioimage.
Here is one from Sightlogix.
I know the Sightlogix can have multiple fixed cameras tied to a single PTZ. I'm pretty sure ioimage can as well.
The Sightlogix uses a positioning scheme on the fixed camera to drive the PTZ. PTZ has no analytics. Not the most accurate, but seems to get the job done.
Ioimage does an actual handoff, where the fixed camera triggers the PTZ to a preset and then the analytics in the PTZ takes over for tracking. This seems to be a bit more accurate.
Both perform best in low traffic areas. The original poster implied he wanted to follow a unique individual across several PTZ cameras. Neither Sightlogix nor Ioimage has the ability to uniquely identify a person, although if they were the only person (ie low traffic) it might work.