Alarming during non-events

[ExplodingLemur]

I'm trying to figure out the black magic that is ZoneMinder motion detection.

We have our cameras set up in an area where the lights are turned off and on regularly. I'm running into several problems:
ZM doesn't always detect motion of people walking through zones, or don't detect until someone is halfway through the frame.
ZM triggers alarms when the lights are turned off and on (though I'm getting closer to resolving this with max alarm area settings, there are still some issues with the cameras alarming several seconds after the lights have come on, on areas where nothing has changed at all)
ZM triggers alarms when the lights are off, due to the cameras dropping shutter speed and increasing exposure time to compensate for low light, which causes more noise in the images.

Are there any general tips I can use to fight these issues? Even reading through the documentation, I'm still not clear on exactly what I should be playing with to try to fix all of these problems.

(ZoneMinder 1.22.3, Debian Etch, downloaded the source package and recompiled it to use in Etch Stable, cameras are Axis 2100 network cameras)

[ExplodingLemur]

Ok, here's an example:

Detection type: blobs
Min/max pixel threshold: 30/0
Filter width/height: 7/7
min/max alarmed area: 7/33
min/max filtered area: 5/0
min/max blob area: 4/0
min/max blobs: 1/0


Detection type: blobs
Min/max pixel threshold: 30/0
Filter width/height: 7/7
min/max alarmed area: 6/33
min/max filtered area: 4/0
min/max blob area: 3/0
min/max blobs: 1/0

I have four event exports that show me walking up the aisle, down the aisle, turning off the lights, and then turning on the lights. If someone could take a look at them and give me some ideas, that would be great.
Walking up the aisle
Walking down the aisle
Turning off the lights
Turning on the lights

Frame stats

[ExplodingLemur]

Here's another image...I was walking down the aisle waving my arms in the air. Note in the image it detects motion where I'm not, rather than where I am. does this have to do with the reference image blend percentage? If so, can someone explain to me how changing that percentage will affect motion detection? I'm not clear on it from the documentation.



As for the actual pixel areas for motion detection, I pulled out my trusty calculator and figured out how many pixels a person is at various distances from our cameras...should be useful when tuning the rest of the cameras, to give me better minimum/maximum blob/alarm areas.