Re: Video capture - multiple sources

[Ant Skelton <ant@xxxxxxx>]

Kevin Hawkins wrote:

> 	Ant
>
> 		- something struck me here - given everyone's obsession with
occupancy
> detection do we have here the bones for a system that uses very cheap
mono
> (L30) cameras in each room and an adaptation of your software that
could
> provide occupancy information on to third party applications ?


I've done a brief test of this using my current motion detector, but the
problem is that a human figure in the field of view which is stationary
for a long period is gradually assimilated into the model for the
background of the scene. The practical upshot if which is, you end up
with pretty much the same problem as with the PIR approach.

This is a known weakness of temporal differencing motion detection - and
more sensitive techniques require exponentially greater processing
power. For example, I swapped my simple stochastic background model for
a Kalman filter estimation, and the frame rate was reduced by a factor
of 10 ;)

The old 233MHz box I used to run all this crap on has finally had an
upgrade though, and I'm working on a new MD algorithm that has a minimal
scene understanding layer that can be fedback to the temporal
differencing component to greatly improve sensitivity at minimal extra
processing cost.

Initial results look promising: it's currently able to spot 'things that
look like cars' and 'things that look like humans', and I'm hoping to be
able to track these elements even if they move into the scene and are
stationary for a long period of time before moving off again.

So, I'll let you know how the ?500 occupancy sensor project goes ;)

cheers

ant