In the afternoon the sun casts a shadow from a tree in my front yard. The movement of the leaves causes many many false triggers. I'm using zones to mask and also Obj Detection which is set at the default. Any suggestions as to how to eliminate the false triggers?
The tree shadow problem is the reason I implemented Deepstack. My CPU works harder, but I now only get alerts for vehicles and people despite moving shadows in the wind.
I almost never use zones in Blue Iris for that reason. I use deep IVS within the cams. I'm impressed by Dahau's own built in AI if you will. That plus DeepStack seems to be the secret sauce for cutting down on false triggers.
In addition to the above advice, you could change make time to 1 second - that is a large field of view and only superman would be in and out under 1 second lol.
The tree shadow problem is the reason I implemented Deepstack. My CPU works harder, but I now only get alerts for vehicles and people despite moving shadows in the wind.
I've just started running DS and still learning about it. Is there an actual GUI interface to see what DS is doing? The only link I saw was browser page which merely says ACTIVE
I almost never use zones in Blue Iris for that reason. I use deep IVS within the cams. I'm impressed by Dahau's own built in AI if you will. That plus DeepStack seems to be the secret sauce for cutting down on false triggers.
I've setup IVS in the cameras using the AI portion of the NVR and have checked that it is setup within the camera's web interface.. The NVR just records everything, but the AI search seems to indicate that IVS is working well.
My life turned around when I started using AI (CPU). Very few false triggers at all from day one of AI - the 'shadow/bunny/breeze triggers' were killing me!! However, I don't need to see the inner workings of DS (though I am technical). Sounds like you want to deploy DS into a container so you can poke and prod it. I try to keep all that mess at work!
I let BI start DS and call DS all from within BI itself. But you can set it up so that BI calls DS from within a container.
In addition to the above advice, you could change make time to 1 second - that is a large field of view and only superman would be in and out under 1 second lol.
After much experimenting (and reading) I've gotten DS implemented and it seems to have significantly cut down on the false triggers. But I still have a couple of questions about DS. Does DS learn on its own over time? When I first installed it it would not recognize a dog then a few hours later it started to recognize the same dog. OTH, DS still thinks that a tree in my front yard is a person and the raised pvc garden in the back yard is confusing DS which thinks it is either a"bench' or a "bus" Will DS ever learn that these objects are not a person,bench or bus. I'm using the custom models "combined" and "dark"
Why not start with the defaults and try that for a week? Using custom models from the start prevents you from knowing what deepstack can do baseline. It gives you behavior to compare to and deepens your knowledge of AI in general.
AI in this implementation does not learn anything. It only goes off the models that are loaded
Whatever model you use, built-in, combined, dark or whatever, specify what objects you want DS to detect . Enter them separated by commas and no spaces. DS is also very case sensitive so they need to be exactly as listed for each specific model.
After much experimenting (and reading) I've gotten DS implemented and it seems to have significantly cut down on the false triggers. But I still have a couple of questions about DS. Does DS learn on its own over time? When I first installed it it would not recognize a dog then a few hours later it started to recognize the same dog. OTH, DS still thinks that a tree in my front yard is a person and the raised pvc garden in the back yard is confusing DS which thinks it is either a"bench' or a "bus" Will DS ever learn that these objects are not a person,bench or bus. I'm using the custom models "combined" and "dark"
The AI was repeatedly identifying a particular tree on my property as a person. This happened at the same time of day when the sunlight struck it at a particular angle. I remedied it by taking that small portion of tree out of the motion Detection zone.
Yeah, I have outdoor furniture on my back patio and during certain times of the year during certain times of the day, it all of a sudden thinks the closest chair is a person and I get an alert (further depends on which direction the chair was left at from its last use).
But again, a few false alerts here and there versus false alerts all week and I am a much happier alert recipient.
Anytime there is movement (like sun through blowing leaves) in the scene and the image is sent to deepstack, you may get the static item box of it is like a car not moving.
For me it is not static as the sun is filtered down through the leaves of a tree and the tree is in a breeze. AI is still very early in its development in general and thinks that dappled sunlight is movement
in the case of the pvc raised bed it was an Alert - DS kept oscillating between bus and bench. I maybe seeing the AI analysis on a pic caused by a tree leaves moving. Another camera which has an oblique view of the garden doesn't have the problem but its FOV doesn't include the trees behind the garden.
Also, I took my busiest camera and turned off Mot Det and AI and setup tripwires in the camera and tested it but it never sent and alert back to BI. I'm resetting the cam back to its original setup. Too many moving parts I need to do more research on all of this. You guys are WAY smarter than me - thanks for the help