False alarms + missed real ones

Support and queries relating to all previous versions of ZoneMinder
User avatar
robi
Posts: 477
Joined: Sat Mar 17, 2007 10:48 am

False alarms + missed real ones

Post by robi »

I am in a really strange situation. I get false alarms triggered by very little changes, and miss events of real alarming situations.

Look at this series:
Image
Image
Image

To avoid this, I made a preclusive area on the pillar.

I am loosing main events:
Image
Obviously this is the frame that alerted the whole thing - and it's way too late, as the guy came to the car, opened it, turned on the lights and started moving afterwards, when it finally alarmed. I lost all the events, I only have the frames from the buffer, but it should have been alarmed way sooner.

I get false alarms again:
Image
I must say that the lamp is NOT flickering so real movenent is not happening.

The settings for the active area:
Image

Any ideas?
v1.25.0 + Ubuntu Linux 12.04 Server
User avatar
robi
Posts: 477
Joined: Sat Mar 17, 2007 10:48 am

Post by robi »

Another example for the settings above:

Image

The guy comes into the picture from right to left. The system recognizes movement only at this moment... way too late.
The movement doesn't even get close to the preclusive zones, so that's not the problem.

I watched in Montage a few hours, and noticed big changes that didn't trigger anything. Like cars passing by, etc...

I believe that my settings are sensitive enough...
v1.25.0 + Ubuntu Linux 12.04 Server
User avatar
zoneminder
Site Admin
Posts: 5215
Joined: Wed Jul 09, 2003 2:07 pm
Location: Bristol, UK
Contact:

Post by zoneminder »

I would suggest turning on stats if you haven't already. Then see what is reported for your first events (in the dark above). Click on the frames link in the events listing to see the stats.
Phil
User avatar
robi
Posts: 477
Joined: Sat Mar 17, 2007 10:48 am

Post by robi »

Stats:

Code: Select all

Zone 	Pixel Diff 	Alarm Px 	Filter Px 	Blob Px 	Blobs 	Blob Sizes 	Alarm Limits 	Score
All 	 33 			4646 (4%) 	3792 (3%) 	3130 (2%) 	1 	2430 (2%) 	82,11-106,146 	2
Note that I didn't have the preclusive zones by that time, and the timestamp was in the zone too (I know that's not the reason but mentioned to be exact).
v1.25.0 + Ubuntu Linux 12.04 Server
User avatar
robi
Posts: 477
Joined: Sat Mar 17, 2007 10:48 am

Post by robi »

Any hints?
v1.25.0 + Ubuntu Linux 12.04 Server
jamescollings
Posts: 59
Joined: Wed Nov 22, 2006 11:26 am
Location: Bucks, UK

Post by jamescollings »

in Another thread of yours, I answered your query about a zone covering 103% of the image. In that, I recommended that you might wish to alter your large zone into a set of smaller rectangular zones. I guess I am going to have to repeat that advice. Especially when I combine it with the following snippet (pinched from an earlier posting that I made.
I recommend that you subdivide your image into a set of rectangular zones. Then you can fine tune things depending upon which zones are triggering for the REAL events and which are triggering for the FALSE events. The first step is to turn on STATS for your events, then you can identify which are real and which are false.

================(snippet starts here)===========
My method of fine tuning zones is this:
Set each zone to be really sensitive (the presets help here). Let the monitor run until you have around 100 events. Log in to the SQL database and dump out the contents of Stats and Events (select * from Events; and select * from Stats;). I then feed these into an Excel spreadsheet, and by using some formulas I can tweak my zone settings to see what the impact would have been of raising thresholds. This way, you can see if you would have missed any REAL events by having less sensitive settings on some zones.

Its just my way... but it seems to work. It also allows me to see if any zones are too large/wrong shape. If there is no way of trimming a zone to ignore the falsies, then I check the images to see WHERE the false alarms eminate from (patch of sun flashing in a particular part of the zone).
================(snippet ends here)===========

Hope this helps?
User avatar
robi
Posts: 477
Joined: Sat Mar 17, 2007 10:48 am

Post by robi »

Still in trouble...

I made several smaller active zones and still not satisfactory results. Same behaviour as above.

Could you share a sample of your excel, with those formulas?
v1.25.0 + Ubuntu Linux 12.04 Server
User avatar
robi
Posts: 477
Joined: Sat Mar 17, 2007 10:48 am

Post by robi »

Okay, following your advice, here are my new settings:
Image

Image

Image

And I still get LOTS of these, missing real events, like turning on the car's lights!
Image

Image

Any help would be very much appreciated. Thanks in advance.
v1.25.0 + Ubuntu Linux 12.04 Server
User avatar
robi
Posts: 477
Joined: Sat Mar 17, 2007 10:48 am

Post by robi »

Another example:
A car is passing by with really strong lights:
Image

Image

And detection occurs (!):
Image
The alarm was not even triggered by the event itself!!! The first two images are taken from the buffer of course, however I would expect that the huge change in the first picture would do some triggering...
(the active zone on the right (63) was a bit taller when this happened, but anyway that's not relevant now)

Stats for this frame:
Image

:( :( :(
v1.25.0 + Ubuntu Linux 12.04 Server
jamescollings
Posts: 59
Joined: Wed Nov 22, 2006 11:26 am
Location: Bucks, UK

Post by jamescollings »

OK.. sorry that i haven't sent you the excel spreadsheet... I will get on to that in the next couple of days (poor excuse, I know, but I will do it).

in the meantime. I notice that you are running this image at 10fps (and then 30fps for an alarm).

Could you just cycle through the frames of an alarm event, and just check the timestamp on each frame? Ideally, you should have to click through at least 10 to get the clock to move from one second to another. If you don't, then I wonder if your system is dropping frames?
Alternatively, you might be running so many frames that the changes between each frame is miniscule, and therefore it isn't triggering an alarm.
I tend to run at 4fps (both in normal mode and alarm mode).

It has to be said.... the only way to Fine-tune your monitor settings is to be in a situation whereby you have TOO MANY events (false alarms). If you are in a position where you are MISSING events (true alarms), then there aren't going to be any statistics recorded from which to work out WHY!

Also, you have your "reference Blend Image" set at 2%. I must admit, to not really understanding the impact that changing this has for alarms, so I have chosen to leave mine at the default of 10, and using the zone configuration to fine tune things.

Just a couple of ideas!
User avatar
robi
Posts: 477
Joined: Sat Mar 17, 2007 10:48 am

Post by robi »

jamescollings wrote:OK.. sorry that i haven't sent you the excel spreadsheet... I will get on to that in the next couple of days (poor excuse, I know, but I will do it).

in the meantime. I notice that you are running this image at 10fps (and then 30fps for an alarm).

Could you just cycle through the frames of an alarm event, and just check the timestamp on each frame? Ideally, you should have to click through at least 10 to get the clock to move from one second to another. If you don't, then I wonder if your system is dropping frames?
Alternatively, you might be running so many frames that the changes between each frame is miniscule, and therefore it isn't triggering an alarm.
I tend to run at 4fps (both in normal mode and alarm mode).

It has to be said.... the only way to Fine-tune your monitor settings is to be in a situation whereby you have TOO MANY events (false alarms). If you are in a position where you are MISSING events (true alarms), then there aren't going to be any statistics recorded from which to work out WHY!

Also, you have your "reference Blend Image" set at 2%. I must admit, to not really understanding the impact that changing this has for alarms, so I have chosen to leave mine at the default of 10, and using the zone configuration to fine tune things.

Just a couple of ideas!
Okay, have a look here:
Image
The maximum frame rate the system reaches per channel is 4.21. I am using a 77-type one-chip card with 4 inputs. This is the actual value both in monitor and alarm state.
The settings 'Maximum FPS' and 'Alarm Maximum FPS' are there, however they don't matter as they set maximum values while the card cannot give more, so they are fine.

I cycled some pictures, the second changes at 4 or 5 frames so I guess that's just fine. The machine load is not more that 0.50 usually.

Regarding the 'Reference Image Blend %ge' - if I set it higher, I hardly get events at all, not to talk about real events. I had it at 7, and now down to 2, when I finally made it get at least some.

Actually I have lots of events, about a hundred a day. But they all are false, and none real. I can look at the live picture, see a car passing by without triggering the alarm! But I get tens of false alarms each hour like the picture above.
v1.25.0 + Ubuntu Linux 12.04 Server
User avatar
robi
Posts: 477
Joined: Sat Mar 17, 2007 10:48 am

Post by robi »

Okay, now I have 648 events a day, almost all false alarms. Real events not catched at all. Please, please help!
v1.25.0 + Ubuntu Linux 12.04 Server
jamescollings
Posts: 59
Joined: Wed Nov 22, 2006 11:26 am
Location: Bucks, UK

Post by jamescollings »

Unfortunately no-one else appears to be jumping in, so you're stuck with me :shock:

until we can catch your "real" events there isn't much we can do about the "falsies".

I had a thought last night... I assume that you have the STATS turned on? If so, you could try triggering an alarm manually, and then acting the part of the "target" yourself. i.e. Trigger an alarm, then walk/drive in front on the camera in the manner that you want to cause a "real" event. Then you can see from the stats what a "REAL" event looks like (in terms of the change in pixel values, % of the zone that is matched etc). It is conceivable that you will be able to filter out the FALSE events by using the "MAX" value (for the alarmed/filtered/blob area).

I have to say... fine tuning your zones to get rid of all of the false events without losing a single REAL event is a long slow process. I spent ages getting this right with my 2 cameras in Winter... and now that it is summer, the sun is coming at a different angle and I'll have to re-do them all again :evil:

Out of interest, what sort of activity would you like to trigger a REAL event, and is there an activity (apart from light flickering) that should never trigger an event? I only ask, since it might prompt me (or others) to recommend an alternative zone setup.

Keep at it... we'll get there in the end!
jameswilson
Posts: 5111
Joined: Wed Jun 08, 2005 8:07 pm
Location: Midlands UK

Post by jameswilson »

setting detection is a trial en error thing and im far from expert at it.
I notice you have some high percentages of screen. Id look at lowering these, starting witha single zone and then increasing sensitivity till you get reliable detection. You will still get false alarms but once we have relaible detection (real and false) then we can work on the settings to remove the false ones
James Wilson

Disclaimer: The above is pure theory and may work on a good day with the wind behind it. etc etc.
http://www.securitywarehouse.co.uk
User avatar
robi
Posts: 477
Joined: Sat Mar 17, 2007 10:48 am

Post by robi »

jamescollings wrote:Unfortunately no-one else appears to be jumping in, so you're stuck with me :shock:

Trigger an alarm, then walk/drive in front on the camera in the manner that you want to cause a "real" event. Then you can see from the stats what a "REAL" event looks like...

Out of interest, what sort of activity would you like to trigger a REAL event, and is there an activity (apart from light flickering) that should never trigger an event? I only ask, since it might prompt me (or others) to recommend an alternative zone setup.

Keep at it... we'll get there in the end!
Well thanks alot for your time.

Stats are on, yes. I'll go down there and do what you said.

The events I'd be interested would be typical: anyone approaching the car you see on the picture. I wouldn't mind falsies like motion around the other car on the right, or headlights reflection from other cars passing by or things like that. But alarms triggered by the camera noise and not seeing real events make me real depressed :(
v1.25.0 + Ubuntu Linux 12.04 Server
Locked