Excessive CPU load with newer versions than 5.8.6.7 version of BI

[Gyula614]

If you go down to D1 on your sub streams you will see a massive improvement. If you are running your substeams at 1MP/2MP then why have substreams turned on at all? Might as well disable them and just use a single stream.

Thanks, will try D1 or 640*480 whichever is available on the camera. full resolution at high fps/bit rate must be used for identification. 3 streams are ideal I think because phone apps can use a low bit-rate low resolution secondary and AI can use a high bit rate fixed res. third.

 

[Gyula614]

Just did an experimental test with CP AI Benchmark tab. Got an image from the web:


Rescaled that to various sizes: 4k, 2k, 1280, 720, 640, 320, 160 wide images and provided them to the AI benchmark.
Then checked the resulting detection %, FPS and CPU load.

• Detection % of the person on the 160 wide was less by 10% (68 %) than on the others. Others (320 ... 4k) had the same 78-79 %).
• FPS from 160 to 720 width was pretty much the same. Above 720 it was in inverse proportion to the width.
• CPU load was around 10 % up to 720. Above 720 it rose to 40 % at 4k. It may be due to CPU-based downscaling to AI input.

From these I think it seems that:

• Anything above 720 width has no advantage (as the CPU downscales anyway, not the AI is that processes...) in detection accuracy.
• Anything above 720 significantly prolonges AI detection time (CPU time may be added) and loads the CPU, hence also requires more power and generates more heat.

It's good to know for AI detection. I wanted to see it for sure.

 

[Gyula614]

However now both YoloV5 and YoloV8 give much more false positive detections, in CP AI 2.6.2 than it was in 2.0.8.
So will revert to 2.0.8 tomorrow. Later version should be better I thought but it seems that it has kind of reached its limits by many reasons.

 

[tofu]

I have the same issue with the same version number of BI. Running latest cpai--

Idle is 5%, but when motion hits I spike to 100%. Issue is not present on 5.8.6.7.

If I turn off ai, the issue is not present.

On 5.8.6.7 and below, motion rarely brings CPU usage above 10-12%, even with AI enabled and functioning correctly.

 

[wpiman]

I noticed my AI MQTT messages changed after an upgrade. I rolled back to 5.8.7.1 and the messages are working as I expect.

Rolled BACK to 5.8.7.1... The message are coming in as intended. Shows me the direction....


5/15/2024 11:31:03 AM
Plate detected: 4986ZH Probability: 98% MOTION TYPE Motion_C>D
5/15/2024 11:23:35 AM
Plate detected: 3KMH56 Probability: 100% MOTION TYPE Motion_D>C

This is my log BEFORE the rollback.... It is reporting motion within the zones....

5/15/2024 9:55:02 AM
Plate detected: 1FCJ27 Probability: 100% MOTION TYPE Motion_ABC
5/15/2024 9:50:22 AM
Plate detected: 4BJG68 Probability: 99% MOTION TYPE Motion_ABC

 

[tofu]

I narrowed the high CPU usage down to when it sends images to the cpai server

Less images are less cpu usage

It has nothing to do with cpai because the issue persists even if I kill the cpai service.

Image burn in, static objects, time between images, "on alert" function -- all had no noticeable effect whether disabled or enabled

This is only a problem post 5.8.6.7

 

[Gyula614]

I narrowed the high CPU usage down to when it sends images to the cpai server

Less images are less cpu usage

It has nothing to do with cpai because the issue persists even if I kill the cpai service.

Image burn in, static objects, time between images, "on alert" function -- all had no noticeable effect whether disabled or enabled

This is only a problem post 5.8.6.7

Great. Do you use blank-out areas?

 

[tofu]

Great. Do you use blank-out areas?

Nope

I suppose the best bet would be to contact the dev, but it's perplexing that this is only happening to a handful of users

 

[Gyula614]

Nope

I suppose the best bet would be to contact the dev, but it's perplexing that this is only happening to a handful of users

That's a good idea. I hope they watch IPCamTalk. It's their forum. Was wondering on other surveillance product but they are expensive like crazy.
My best feature would be thrown object detection. But as long as the AI doesn't keep track of the identified objects on the sequential images it's not able to do it.
Also there are many false detections even on a single picture. To reduce that I think it would be beneficial if the AI would return the probability of other categories of an object. Like if it says something is a dog by 75 % and also detects as a person by 80 % then the thresholds could be tuned for the multiple categories to get more accurate results. Now it detects our dog as a person many times. I could simpily disable that by setting if it's also a dog by more than 70 % then it's not a person. But it narrows down to object handling as it should say some %s from the same group of pixels (object) that it also can't do right now. It just says something from the image now, from somewhere, about some pixels... so pretty limited. Because now, if there's a dog beside a man, I just can't enumerate anything on the exclude list in order not to disable the person detection...

 

[wittaj]

That's a good idea. I hope they watch IPCamTalk. It's their forum. Was wondering on other surveillance product but they are expensive like crazy.
My best feature would be thrown object detection. But as long as the AI doesn't keep track of the identified objects on the sequential images it's not able to do it.
Also there are many false detections even on a single picture. To reduce that I think it would be beneficial if the AI would return the probability of other categories of an object. Like if it says something is a dog by 75 % and also detects as a person by 80 % then the thresholds could be tuned for the multiple categories to get more accurate results. Now it detects our dog as a person many times. I could simpily disable that by setting if it's also a dog by more than 70 % then it's not a person. But it narrows down to object handling as it should say some %s from the same group of pixels (object) that it also can't do right now. It just says something from the image now, from somewhere, about some pixels... so pretty limited. Because now, if there's a dog beside a man, I just can't enumerate anything on the exclude list in order not to disable the person detection...

IPCamTalk is not affiliated with BI, nor do they view this site.

 

[Binod Bhandari]

Yeah, it's not normal for CPU usage to spike like that with updates. Rolling back to an earlier version could be a quick fix until they sort it out.