Beginning with AI

[Kazzaw]

Hi all,

I've been reading through numerous threads over the last few months, but the continuous updates and feature changes to both CPAI and BlueIris are starting to get a little overwhelming.

I've been using BI successfully now for 6 months, and I'm starting to look into advanced features of AI - Including person detection, animal detection, LPR from a single camera and Face Detection.

My hardware is as follows (taken from the CPAI screen)

Server version:   2.5.6
System:           Windows
Operating System: Windows (Microsoft Windows 10.0.17763)
CPUs:             12th Gen Intel(R) Core(TM) i5-12600 (Intel)
                  1 CPU x 6 cores. 12 logical processors (x64)
GPU (Primary):    Quadro P600 (2 GiB) (NVIDIA)
                  Driver: 537.70, CUDA: 12.2 (up to: 12.2), Compute: 6.1, cuDNN: 8.5
System RAM:       31 GiB
Platform:         Windows
BuildConfig:      Release
Execution Env:    Native
Runtime Env:      Production
Runtimes installed:
  .NET runtime:     8.0.0
  .NET SDK:         8.0.100
  Default Python:   Not found
  Go:               Not found
  NodeJS:           Not found
Video adapter info:
  NVIDIA Quadro P600:
    Driver Version     31.0.15.3770
    Video Processor    Quadro P600
System GPU info:
  GPU 3D Usage       3%
  GPU RAM Usage      1.9 GiB
Global Environment variables:
  CPAI_APPROOTPATH = <root>
  CPAI_PORT        = 32168

It's important to note that I do not have access to the iGPU on this machine, as it is virtualized and something else uses the iGPU. I do have performance issues, and if I can't rectify them fully (which I think may be the cause of some of my issues) I'll look to move the BI/CPAI setup to a dedicated host.

For now, I have 3 cameras. Front North, Front South and South 180.

Front North & South usually get triggered simultaneously as they both face the road. I do want this, and I want the alerts to be labelled as to what was detected, as this makes it easier when searching for street traffic and related incidents. These cameras each have 3 different Zones configured. Zone A, which is street traffic (Which I need to confirm with AI, but I don't have notifications enabled in this zone), Zone B which is the Driveway, and Zone G which is the yard. Zone B and G will give different alerts (critical or normal) as per the time of day/schedule configured (and working well, with automations with home and away configured through HomeAssistant).

I've dabbled in the custom AI models, but I also think the amount of models I am running the images through is causing bottlenecks as well, and I'm struggling to figure out how to use my alert/ai with the models, when to include the ':0' feature, and I'm yet to have much luck with plates being captured.

I don't want to load up this post with too much information, and would rather go as I change things, and find more helpful resources (that are current would be ideal!)

My first target is going to be getting my zones, AI confirmation and alerting working properly. I've attached some photos from Camera Front South, Showing AI, Motion settings, Zone settings and BI AI Settings - These should all be named accordingly.

If I can get this one alerting properly, hopefully I can do the rest myself without having to lean on the community so much.

Can anyone please look at these settings and let me know if it looks OK? I'm happy to provide anything further.

 

[VideoDad]

First, it feels like you are trying to do too much with a single camera. Pick a single purpose and stick with that. You can't expect that angle to get much in the way of detail on the street or the driveway and may be looking down too much to get more than people's heads if they were to walk across your lawn.

As for your settings, you have the default object model still enabled. If you are planning to use custom models (which I recommend because they only contain objects relevant to a security camera and are faster), then on the global AI settings, uncheck "Default object detection". You seem to be also considering facial recognition which I wouldn't do with this view, so no need to save unrecognized faces.

Next, you might want to simplify your motion triggering. It seems like you've set it to only detect if the motion covers a large area of the screen. You probably want that smaller, not bigger because many of the objects in that particular view will be smaller.

Regarding zones, I've had problems trying to do different detections in multiple zones. So again, consider if you aren't doing too much with a single camera. As for the specific AI settings, you've picked the ipcam-combined custom model which gives you the most choice for objects. I see you've picked 'person,car,truck,cat,dog,bird'. I guess you aren't interested in motorcycles, buses or bicycles? Do realize that sometimes a big truck might get detected as a bus, so you might want that label too.

From that camera view I don't know how useful the animal detection will be, but nevermind that. In terms of the 'Mark as vehicle' choices, you have 'vehicle' which isn't even a detected object name in the model. You only want the actual labels from your 'To confirm' list that are actual vehicles. In your case, only "car,truck" at the moment, though if you do add bus, then you already have it here.

I see you have .DAT files being generated. That will be helpful to see what is actually being detected. I would definitely monitor any triggered clips over the next few days and look at the resulting AI processing on the .DAT file. That will tell you a lot about what is/isn't being detected, if you have extraneous models running on a camera, if things are taking too long, if objects are being missed, etc.

 

[Kazzaw]

As for your settings, you have the default object model still enabled. If you are planning to use custom models (which I recommend because they only contain objects relevant to a security camera and are faster), then on the global AI settings, uncheck "Default object detection".

Thanks, I've disabled the default object model now

You seem to be also considering facial recognition which I wouldn't do with this view, so no need to save unrecognized faces.

It seems to perform OK on this camera, although it hasn't been working properly lately unless on reboot. I recall reading another thread that mentions that face recognition is a bit hit and miss, so I'm not too concerned at the moment.

Next, you might want to simplify your motion triggering. It seems like you've set it to only detect if the motion covers a large area of the screen.

I'm not too sure how I can bring this down - I definitely need to capture the road, as this allows me to flag when vehicles drive past (has been helpful in the past when police requested some evidence, I was able to navigate to a vehicle event rather than browse through an hours worth of footage). Perhaps I could get rid of the driveway zone as its mainly monitored by the other camera.

I see you've picked 'person,car,truck,cat,dog,bird'. I guess you aren't interested in motorcycles, buses or bicycles? Do realize that sometimes a big truck might get detected as a bus, so you might want that label too. From that camera view I don't know how useful the animal detection will be, but nevermind that.

Don't need bird, I think that was from when I was playing around. Dog and Cat work well though as we have quite a few strays in the neighbourhood

I see you have .DAT files being generated. That will be helpful to see what is actually being detected. I would definitely monitor any triggered clips over the next few days and look at the resulting AI processing on the .DAT file. That will tell you a lot about what is/isn't being detected, if you have extraneous models running on a camera, if things are taking too long, if objects are being missed, etc.

Thanks, the .DAT files have been super helpful, and I assume they will be more so now that I actually have the AI configured a little better. I'll monitor over the next while and see how I go. Appreciate your time

 

[Kazzaw]

Back again - I've been going well with AI on that specific camera, however I'm now struggling with both my front cameras detecting my car and notifying (critical IOS alert) of its presence at night time.

I've tried to configure static object detection, but I just can't seem to get it to work. Any suggestions on this one? I'm guessing its being triggered by the headlights in the camera, which is understandable

 

[RogMagog]

For your setup, it might help to start with simplifying the number of models or zones you’re testing. Sometimes, less is more, especially if you’re seeing bottlenecks. I had success using fewer, more focused models, which helped with performance and accuracy.

As for alerts, make sure your zones are set up with distinct triggers and notifications. For instance, separating critical and normal alerts based on time of day worked well for me in a similar system.
If you’re looking for the best ai for business, trying out different configurations and tools might also help fine-tune your setup.