Hi, good job on FF2.0
I was trying the FFComputeMeanValues on 4 rectangle snapshots on a webcam pic to determine motion detection and specifically to understand which direction the person is going.
I use this rig
so I put rectangles on door A, B, C and D (the white rectangle in bottomright)
I get the snapshots every 100ms and I collect every subsequent detection area in a string with all the areas interested, so, in theory, analyzing the first and the last detection I should know the starting and ending crossing and so the direction of the person.
It works quite well, but I get some problems:
1) At the right of D there is a window so when someone passes in that area, can cast some shadows on door A and this fires up the MeanValues of A without someone crosses there. (A false detection)
2) Apart that, sometimes if someone crosses from D to A or viceversa, I get sequences like this:
from DtoA: DBDBDABAC
from AtoD: ABDDADA
that means in the final motion there is still a motion in an area different from the real final.
For C area it could be some pixels are still out of range after passing through A.
For the A area, it could be the door is still slowly moving when arriving to D
The point is that these false detection could really be true if someone changes idea and wants really to go to C or to return to A from where it was coming...
So, how could I get the real direction of a person avoiding false detection?
Edited by frank10, 29 April 2012 - 04:13 PM.