 |
People counting
Sebastian Johnson from Sweden [1 posts] |
14 year
|
A few months ago we had some discussion here in the forum regarding the possibility to accurately count people moving in and out beneath a ceiling mounted camera. I sent some actual videos from a test setup in a shop as well as some images. You had some ideas around this, but i never got any final words from You in the matter. Was it possible to track and count the people in the videos i sent with any accuracy?
Kind regards
Sebastian Johnson
Sweden
|
|
|
|
Anonymous |
14 year
|
Sebastian,
We've actually have been continuing to work on this issue. The launch of the Blob Tracking module and the various improvements to the Movement module were as a result of trying to increase the accuracy of the counter using your images. This has improved things somewhat but not to a high enough degree to be usable. The main issue has been the grated floor changing intensity when two people are close to each other as they then appear to merge with respect to movement and count as one. Thus a possibility is to change the floor to remove the grating but we assumed this is not a possibility.
Another possibility is to zoom the camera into the entrance/exit a little more in order to create more pixel space between people. This increase in resolution should be better handled in the movement module.
Another idea is to create a very top down view that always sees people as perpendicular as possible from above. The reason for this is you could then just count the number of pixels and divide by what is considered a normal amount for a single person. This technique has been used in other people counter systems and appears to be more robust than expected. The main accuracy relies on the camera not warping the image. In your case people at the center of the image appear larger due to more of their backside being visible. Towards the lower edges they appear in a more ideal solution in a straight top/down view. Perhaps if possible just placing the camera more center just before/after the exit would allow this technique to work much more reliably.
Naturally these suggestions is just an idea and may or may not improve the situation. One day software should be able to process your images with very high accuracy but that might not be for a while. If you need this as a solution ASAP then changing the camera view is probably the fastest solution. We will, however, continue to work/research on this issue as it is a challenging problem!
Also, how accurate do you need the counter to be? Current IR/beam counters will often mistake more than one person for a single person so perhaps the visual tracker should at least be as good as that?
STeven.
|
|
|
|
Andrew from Australia [6 posts] |
14 year
|
I have been looking at people counting over the last day or so using the video Sebastian supplied. I have come up with the robo file attached. This examines a small area in front of the door where everyone passes, equalises the colours and intensity, then uses the movement module to find the moving people. All this works pretty well.
The next stage I shall use will be to pass the image on to MATLAB. I am pretty confident I can write a MATLAB script to differentiate the people and work out whether they are coming or going by looking at the time history of the final image. However I do not know how to use RoboRealm to do the same task. Maybe once I do it in MATLAB I can make a VB script to do something similar, but creating the VB script directly from RR is just too hard.
 program.robo
|
|
 |
Agustin Gudino from Spain [2 posts] |
13 year
|
|
Well, I am new to RoboRealm and I just testing it... About people counting I think the best option is using the Deph Kinect Camera with some filtering and matching I think. This is what I am researching now.
|
|
|
|
Anonymous |
13 year
|
Agustin,
Yes, that's a good use of the Kinect. If you can mount the Kinect looking stright down at the floor people will appear as large peaks in the depth map which can be threshold, isolated and probably fed into the Blob Tracker for counting. You will still have some issues if people are very close to each other but that will be at a minimal.
Let us know how it goes. You may need a relative thresholding in order to isolate peoples heads at different heights (i.e. identify all peaks of at least X cm in depth). This is not unlike detecting a local maximum. If you run into this issue post an image and we'll see what we can use for local peak detection.
STeven.
|
|
This Thread