BI setup and image detection

[kolbasz]

When I first started to set things up, I jumped right to the 2 camera AI trigger setup in BI.

That is, I set up the AI tool with deepstack image detection, and from there, triggered the second camera with higher resolution recording. However, I have long thought about this setup and thought maybe I need to slow down and walk before I run.

What is the typical setup for people? Do most people just configure BI, set the detection area and just let BI handle the triggering of the recording? Is the AI tool an over complexity or is it something that a lot of people are using? I assume there is an ask for it as development continues and I see improvements with it along the way.

I currently run 2 cameras, one is a doorbell camera, and the other a 5542T-ZE. I want to understand how things work in BI and in general and then get things tweaked from there.

1. when BI makes a detection, is it BI that actually detects motion or does BI get a trigger from the camera feed? Such as, my doorbell camera sends me motion alerts, are those what also trigger BI to capture an image jpg, or is BI acting independently?
2. Is the AI tool something that I want as part of my setup?
3. When capturing 24/7, what do people typically choose as the recording resolution? Is the 24/7 feed pulled from the secondary feed or is it just saved as a lower quality, is that an option?
3a. I want to make sure I understand the use of the secondary feed as I know when I started it was said the secondary feed in BI uses very low CPU etc, and want to make sure I properly understand the benefits and uses of it.
4. detection zones, are these more for better detecting or reducing false detections in areas we do not worry about as much for alerts, like the middle of the yard where kids play as this creates lots of recordings we may not really need.

things have been OK to date, but I also know when I first got things installed, I was a bit excited with it all and just sort of threw it all together so it worked. I want to now go back and finalize things and improve on what I have, so am looking for some ideas and pointers.

 

[sebastiantombs]

Simplest way is to record 24/7, use BI or camera ONVIF for motion detection and never miss anything. Disk space is cheap unless you want months of recorded data or are trying to produce an epic with high frame and bit rates.

 

[kolbasz]

what about quality? I know we call these 4k and even 8k cameras but is that resolution even required when it hits the disk? Does it make sense to record at 1080p? As you note, I am not making an epic of package deliveries.

when you mention using BI or camera ONVIF for detection, is this to imply the BI detection is independent of the camera itself and it is merely the software detecting changes to the image/feed? just confirming this point for myself

 

[wittaj]

You can have the camera trigger motion for BI and then you do not use BI for motion detection and BI simply records based on receiving the motion trigger. It can bring the BI CPU down as well.

If your camera has AI built in then you do not need AI Tools/Deepstack and use the camera AI as it has been proven to be more accurate than Deepstack (although you can only AI for a car or person with camera AI, so if you want it to AI for bears and dogs, etc., then you need a 3rd party).

Depending on your situation and what you are trying to accomplish, you may not even AI Tools. If you can set up motion detection correctly with IVS rules and or motion settings in BI to be fairly accurate, then Deepstack is just a resource hog on your system...

We shouldn't even be purchasing 4k and 8k cameras as they are on the same sensors as 2MP and 4MP cameras, so they need A TON more light. Down resolution of a 4K camera to 2MP doesn't change the physics fact that the sensor is the same size - you would be better off purchasing the 2MP camera.

If your monitor cannot display more than 1080P, then it probably is a waste of space to be recording 8K...

 

[sebastiantombs]

BI records at both full resolution and at sub stream resolution, assuming you're using sub streams, since version 5.2.7.x I think. I know it does since 5.3.x.x. Generally, if a video needs to be preserved, one would want the original, native, resolution to preserve as much detail as possible. That can be aa key factor if the video is needed or used as evidence by the authorities.

BI does motion detection based on scene changes and allows control of size and contrast to determine that motion. It also has a few different algorithms that add more flexibility. Dahua, Hikvision and some other decent brand cameras allow setting up "trip lines", "intrusion boxes", regular motion detection and some even offer basic AI for people and vehicle detection.

Both BI and camera motion detection have their uses. I use a mix of both sometimes to limit false detections in BI from moving shadows and in other cases to detect movement, very reliably, at distances that BI motion detection struggle with.

 

[wittaj]

Another benefit the substream option has about bringing down the CPU usage that most do not talk about or realize is during playback. Without the substream option, if you playback all the cameras at once (like you want to see a person go from the front to the side to the back of the house for example), then all of the cameras on the screen are playing back at mainstream resolution and that will peg the CPU hard and if you have enough cameras, it will max out the CPU. With the substream option, all the cameras in a tile/cascade show substream until you select one camera to view on the screen, then that camera only goes to mainstream.

 

[kolbasz]

BI records at both full resolution and at sub stream resolution, assuming you're using sub streams, since version 5.2.7.x I think. I know it does since 5.3.x.x. Generally, if a video needs to be preserved, one would want the original, native, resolution to preserve as much detail as possible. That can be aa key factor if the video is needed or used as evidence by the authorities.

BI does motion detection based on scene changes and allows control of size and contrast to determine that motion. It also has a few different algorithms that add more flexibility. Dahua, Hikvision and some other decent brand cameras allow setting up "trip lines", "intrusion boxes", regular motion detection and some even offer basic AI for people and vehicle detection.

Both BI and camera motion detection have their uses. I use a mix of both sometimes to limit false detections in BI from moving shadows and in other cases to detect movement, very reliably, at distances that BI motion detection struggle with.

Found a substream2 option, don't remember it originally. It allows me to set image at 1080p, 720,etc. I think this should get me past the problem with the substream2 quality being horrible. Need to test with this and also mess with the detection zones, both camera and bi.

I recall looking at it in the camera GUI, but not sure if/how it triggers the alternate feed recording.

 

[wittaj]

Found a substream2 option, don't remember it originally. It allows me to set image at 1080p, 720,etc. I think this should get me past the problem with the substream2 quality being horrible. Need to test with this and also mess with the detection zones, both camera and bi.

I recall looking at it in the camera GUI, but not sure if/how it triggers the alternate feed recording.

True, but if the resolution is too high, it kinda defeats the purpose of the substream... I run all of mine at D1 and 256bitrate and it looks fine. Maybe a smidge not clear, but not enough that it is worth bumping the resolution and bitrate as that is what going to mainstream is for.

 

[kolbasz]

perhaps I need to try it again. It was unviewable previously and everything was grainy, etc, to the point where it made no sense to even record 24/7 in that format. I know too I had lots of issues with vbr to cbr as the vbr is notoriously bad on the 5442 it seems
.

 

[sebastiantombs]

I run the sub streams at the lowest resolution that is allowed by the camera. When viewing a clip or the camera as a single view the full resolution comes up. Maybe my eyes aren't what they used to be, but I can see the difference in a multi-camera view and it is not as sharp, but that's not all that critical to me. Keeping the CPU utilization low it my goal.

 

[samplenhold]

8K? That's 33MP. I don't know of any cam that is 8K.

 

[kolbasz]

8K? That's 33MP. I don't know of any cam that is 8K.

sorry, 4k 8MP is more like it.

 

[kolbasz]

I run the sub streams at the lowest resolution that is allowed by the camera. When viewing a clip or the camera as a single view the full resolution comes up. Maybe my eyes aren't what they used to be, but I can see the difference in a multi-camera view and it is not as sharp, but that's not all that critical to me. Keeping the CPU utilization low it my goal.

I get it, but there is the flip side. What happens if detection fails and BI does not flip to higher quality on motion. Now you have low res footage of the event. Is this good enough, I am not sure. I know from my basic tests I was disappointed, but that could be simple user error. I suppose it is a science of finding the happy middle ground that is low CPU and acceptable sub stream quality

 

[wittaj]

BI records both streams, so you can watch playback in mainstream without a trigger.

 

[kolbasz]

BI records both streams, so you can watch playback in mainstream without a trigger.

so I guess I should asks as maybe I am missing the strategy.

Is the strategy to always record the main stream and the point of the sub stream is simply for screen displaying?

My thought was that regardless of having a display or not, that I 24/7 record the substream and then trigger the main stream on motion. However, I am getting the idea that folks prefer to 24/7 record the main stream too? and then just trigger separate event recordings in BI?

 

[Holbs]

I get it, but there is the flip side. What happens if detection fails and BI does not flip to higher quality on motion. Now you have low res footage of the event. Is this good enough, I am not sure. I know from my basic tests I was disappointed, but that could be simple user error. I suppose it is a science of finding the happy middle ground that is low CPU and acceptable sub stream quality

I have 5+ cameras along the front of my house. I rely on the 4 x 5442 series AI 100% (so far, 100% accurate as well) for HUMAN using IVS tripwires/instrusion boxes. I do not rely on just motion detection alone.
Due to some criminal events around my place, I too had to make that decision to decide if some cameras should have a non-blurry non-pixelated substream view so went with D1 and high bitrate on cameras I considered critical. Other cameras, lowest bit rate possible for substreams. I could not afford that 1 out of 100 chance that tripwires alone would work so left one of my 5442's recording 24/7 in hi-res, the other 3 recording high bitrate substreams (which are still substantially smaller K/ps).
Sorry if that sounds confusing. Basically:
1 x 5442 series camera recording 24/7 in high resolution using AI human detection to trigger alerts.
3 x 5442 series cameras running D1 high bitrate using AI human detection to trigger high resolution alerts/recording.
1 x PTZ (still fine tuning).
various older models without AI (5231's) that will have to rely on Deepstack / OnGuard.
You have to decide how trustworthy AI is on specific cameras. I like backups / covering my arse so that is why I do what I do. Watched Terminator too many times to 100% rely soley on AI

 

[wittaj]

Two ways to go about it. One is to 24/7 just the substream and then record triggers on mainstream. Depends on how long you want to retain and how much storage you have.

Most of us probably record the mainstream 24/7, but utilize the substream option to significantly bring down the CPU usage. Not only does it bring down the CPU usage in general, but then on playback of watching multiple cameras at once, it doesn't max out the CPU because the cascade tile of cameras are displaying substream until you select one camera.

I personally decided to record mainstream 24/7 because if something happens, inevitably the pre and post record times are not long enough and it has been sufficient for me to see the going ons of a perp door checking (obviously not clearly at a distance) but certainly enough for observation that wouldn't happen from a substream.

 

[kolbasz]

Two ways to go about it. One is to 24/7 just the substream and then record triggers on mainstream. Depends on how long you want to retain and how much storage you have.

Most of us probably record the mainstream 24/7, but utilize the substream option to significantly bring down the CPU usage. Not only does it bring down the CPU usage in general, but then on playback of watching multiple cameras at once, it doesn't max out the CPU because the cascade tile of cameras are displaying substream until you select one camera.

I personally decided to record mainstream 24/7 because if something happens, inevitably the pre and post record times are not long enough and it has been sufficient for me to see the going ons of a perp door checking (obviously not clearly at a distance) but certainly enough for observation that wouldn't happen from a substream.

curious, what is your storage capacity and retention time? maybe my thoughts are skewed

 

[kolbasz]

I have 5+ cameras along the front of my house. I rely on the 4 x 5442 series AI 100% (so far, 100% accurate as well) for HUMAN using IVS tripwires/instrusion boxes. I do not rely on just motion detection alone.
Due to some criminal events around my place, I too had to make that decision to decide if some cameras should have a non-blurry non-pixelated substream view so went with D1 and high bitrate on cameras I considered critical. Other cameras, lowest bit rate possible for substreams. I could not afford that 1 out of 100 chance that tripwires alone would work so left one of my 5442's recording 24/7 in hi-res, the other 3 recording high bitrate substreams (which are still substantially smaller K/ps).
Sorry if that sounds confusing. Basically:
1 x 5442 series camera recording 24/7 in high resolution using AI human detection to trigger alerts.
3 x 5442 series cameras running D1 high bitrate using AI human detection to trigger high resolution alerts/recording.
1 x PTZ (still fine tuning).
various older models without AI (5231's) that will have to rely on Deepstack / OnGuard.
You have to decide how trustworthy AI is on specific cameras. I like backups / covering my arse so that is why I do what I do. Watched Terminator too many times to 100% rely soley on AI

I think it makes sense. one camera as a catch all just in case. for the rest, rely on the camera technology to switch to the better quality. In the event of a failure, you still have the one stream to fall back on which is a solid strategy.

 

[wittaj]

It comes down to how long do you want to save. I used to be of the mindset of wanting a long retention and would just record on motion, but then things were missed. Then I thought oh I will just record some on motion and the rest 24/7. But many here have said if it happened more than X weeks ago, it probably doesn't matter at that point.

There is some truth to that and if you have the system set up well to cover your own house or property, you should be getting notifications or you see it when you scrub through the video. Then if you notice anything suspicious, just pull or flag that video to save it for later if some neighbor said something happened. I know that if a neighbor said something happened about a month ago and did I see anything, I am not going to spend a lot of time looking at multiple days and times. If they can say a date and a time range, then ok.

I also run LPR, so I have a long history of plates coming through if something happened. I also save the snapshots on motion trigger for a longer period, so I do have snapshots of cars and people walking past my house on the street or sidewalk.