Unreliable IVS Intrusion setup

[mooch91]

Hi all,

I've been setting up my Dahua cameras for IVS Intrusion events. My goal is to try to detect when objects get close to windows and doors within view of my cameras. One is working great so far, the other, not so much.

I've been trying some variations of this setup, a rectangle around the paver area leading up to the door. Sorry for the night pic, the green rectangle represents the Intrusion area, the quarter circle at the top of the Intrusion area is the paver area, and to the left of that is the door.



This setup seems to work well when someone comes in from the right side, but not when someone comes out from the door on the left or in towards the door from the driveway at the top.

Any suggestions for setting this up to have it work more reliably?

Thanks!

 

[wittaj]

A couple of issues.

The camera needs to first see the object and decide if it meets the criteria, so the from the left will be problematic as there is a door that swings open, it is on the edge of the field of view, the box is too small for people coming from the left, the box looks more like around the garbage cans, and the bush is messing with the exposure.

It just comes down to a poor field of view and a bush washing out the image which then make AI harder.

You can continue with tweaks, but trimming the bush would help a lot.

 

[mooch91]

A couple of issues.

The camera needs to first see the object and decide if it meets the criteria, so the from the left will be problematic as there is a door that swings open, it is on the edge of the field of view, the box is too small for people coming from the left, the box looks more like around the garbage cans, and the bush is messing with the exposure.

It just comes down to a poor field of view and a bush washing out the image which then make AI harder.

You can continue with tweaks, but trimming the bush would help a lot.

Would you think that bush should cause such a big issue in the daylight? Because the detection is equally unreliable during the day.

The door swings in to the house, so the first thing that will 'appear' when someone comes out is the person. They usually make a quick left and head on to the driveway.

Are you suggesting expanding the box "in to" the siding of the house and perhaps out further on the driveway?

Thanks!

 

[kolbasz]

I’m just exploring Ivs over blue deep stack, but if the door is above the trash can, I say make the detection box larger.

 

[wittaj]

Yeah, the box is way too small, especially if the door is on the corner "above" the trash cans. I would extend it to the basketball hoop so the camera has time to recognize the person. By the time the camera does now, the person is out of the box.

Further, the trash cans would basically block the lower part of the body, so AI will be problematic with that too.

It is just a tough field of view for that location and what you are trying to accomplish.

 

[looney2ns]

This is why you test proposed cam locations prior to final install, to avoid these issues.

 

[mooch91]

Attached is a daytime shot of the IVS setup so you all can see it a bit better. I expanded the size of the zone. It's picking up cats and groundhogs traveling behind the trash cans quite well, day or night. It's still not getting me coming out the door, nor me coming in from the "top" from the driveway.

 

[mooch91]

For perspective, attached is a shot of me coming out of the door, facing to go to the driveway. I would surely think this would pick me up, especially because it's getting small animals. Maybe it's a matter of timing - I don't know how long Intrusion takes to register, but I can usually get out the door and be around the corner on the driveway in 4-5 seconds I imagine.

 

[sebastiantombs]

Once again, the camera needs to see the object enter the box to trigger the detection algorithm. The box being on the house, on the left side does not allow that to happen which is why you're not detected as you exit. You just appear inside the box. That box needs to move off of the house, onto the ground and away from the house.

Additionally, I would move the camera view more to the right and get the wall of the house out of the frame. This will reduce IR reflection and improve the night video. The bush in the foreground is another problem because it reflects so much IR at night.

 

[wittaj]

And as I said earlier and this picture confirms, it only sees half a human and the AI is looking for a whole body not the upper half.

 

[mooch91]

And as I said earlier and this picture confirms, it only sees half a human and the AI is looking for a whole body not the upper half.

OK, I thought the Intrusion AI just looked for an object of certain size and not a human shape. One of my other cameras has actual Human detection. And I thought checking Appears would allow it to pick up when something just appears in the zone.

 

[wittaj]

You will have to assign it a min size and then it might puck it up.

 

[looney2ns]

OK, I thought the Intrusion AI just looked for an object of certain size and not a human shape. One of my other cameras has actual Human detection. And I thought checking Appears would allow it to pick up when something just appears in the zone.

If you have an SD card in the cam, you can see exactly what the AI is doing with recordings.

 

[looney2ns]

Try something like this, if that doesn't work in the same area, place a tripwire instead with it zigzagged and cover the same approximate area.

 

[tmxv4128]

I also think having seeing more of a body would make it easier to identify. Any chance of moving the trash can out of the way to see if that helps? When looking at something like this, I often have to think in 3D, rather than 2D.