Canceled Alert with 91% deepstack

Post Reply
mdmike41
Posts: 1
Joined: Thu Sep 16, 2021 3:42 pm

Canceled Alert with 91% deepstack

Post by mdmike41 »

I am confused as to why sometimes I do not get alerts when deepstack finds objects. I cant find any rhythm to it, its very intermittent. I got an alert when this car left, but not when it came back an hour later. Deepstack obviously detected it as a car but in BlueIris it said nothing found. Any ideas?
Capture.JPG
Capture.JPG (223.3 KiB) Viewed 6265 times
Alex
Posts: 7
Joined: Fri Aug 27, 2021 7:29 am

Re: Canceled Alert with 91% deepstack

Post by Alex »

I have had the same kind of problems with some of my cameras.
The way I troubbleshooted it was to pull up the deepstack status on one part of the screen and BI on the other.
Then hold ctrl key and doubbleclick on the recording where nothing was found.
The clip now starts at the time of motion instead of the "pre time" and you can see in the deepstack status what deepstack percieved.
Then depending on how many pictures and time between you have set for that specific camera you can step forward some frames and see what deepstack sees.

Example from my end.
In the evening my camera triggered already when lights from the car was shining on the driveway and since I had like 4 pictures every 250ms it meant that
deepstack had already stopped it´s analysis before the car was visable in the frame.
dz93
Posts: 2
Joined: Mon Sep 20, 2021 5:27 pm

Re: Canceled Alert with 91% deepstack

Post by dz93 »

mdmike41 wrote: Thu Sep 16, 2021 3:50 pm I am confused as to why sometimes I do not get alerts when deepstack finds objects. I cant find any rhythm to it, its very intermittent. I got an alert when this car left, but not when it came back an hour later. Deepstack obviously detected it as a car but in BlueIris it said nothing found. Any ideas?

Capture.JPG
On your camera settings, go to the trigger tab and on the artificial intelligence settings where deepstack is located, uncheck the "use mainstream" setting. There's presumably a glitch in blue iris where when using main stream video to feed the AI it doesn't actually send the correct amount of images or is doing it too slowly and results in missed detections. Disabling the use mainstream option will instead use the sub stream feed for deepstack analysis and can send the images to deepstack much faster while also sending all of the set images. You can change how many images are sent by changing the number under "real time images". I usually set mine to 10 every 1second or less if need be.
User avatar
YrbkMgr
Posts: 587
Joined: Sun Nov 24, 2019 12:56 am
Location: Chicagoland

Re: Canceled Alert with 91% deepstack

Post by YrbkMgr »

This is how I think about DeepStack and BI and how I approached the same issue - I'm not saying I'm right, I'm saying that this is how it seems to me...

Regarding the car in the driveway, you should be aware that DeepStack has an option to determine if there are stationary objects and exclude them from the analysis. When enabled, it will cancel an alert for... well... stationary objects. That setting is found in Camera Settings -> Trigger -> Artificial Intelligence. So above all else, if that feature is enabled, you won't see an alert for a parked car even when it's confidence is above the minimum. I'm not saying that's the issue for you, but you should know about it.

From what I've surmised, there are two ways to view a DeepStack analysis: Review the DAT or Analyze the BVR. Each deliver different information.

Analyzing the BVR: Under Testing & Tuning -> Analyze with DeepStack, you are evaluating DS detection performance on a previous recording, period. It delivers no motion detection or trigger information - no "event" information. It has nothing to do with your camera settings. Instead of feeding DS live video, you're feeding it a recording. As far as I can tell, the biggest value is in making sure that your global AI settings perform as expected (face rec, lic plates, etc) - that's why this is under testing and tuning... so if Analyze with DeepStack is detecting everything you expect on a previously recorded video, then that's not where you should be looking.

Reviewing DS Tab of Status Window: DAT Status is the log where you review what happened that caused a "Go/No Go" alert event with the DAT file. DS is the gatekeeper for alerts - camera still triggers and records according to camera settings... alerts are confirmed or cancelled by DS instead of whenever the camera triggers.

In order to analyze what happened (DAT) the first requirement is to enable "Save DeepStack analysis details" (Camera Settings -> Trigger -> Artificial Intelligence). This creates a log of sorts so that one can review when DS began and what it found. It is stored in your Alerts folder by default. From that point on you will be able to view a clip and evaluate what DS detected.

To review the analysis, first open the Status Window and have the Deepstack Tab selected. Enable "Highlight Motion Rectangles" in that DS Tab. Now you're ready. Ctrl-DBL Click on a clip in the Clips pane. Two things happen: first, the clip opens at the frame that triggered the camera, and pauses. The second thing that will happen is the DS Tab of the Status Window will populate with information contained in the DAT file.

DeepStack "log" (dat) :
Deepstack Analysys 6.JPG
Deepstack Analysys 6.JPG (114.52 KiB) Viewed 5855 times
Analyze with Deepstack (bvr):
Deepstack Analysys 7.jpg
Deepstack Analysys 7.jpg (123.31 KiB) Viewed 5855 times

The DAT information indicates if DS motion rectangles line up with the frame that triggered the motion detector. Each timestamp on the tree shows what DS saw at that moment. As far as I can tell, this information tells you how well optimized the overall system is. CPU, GPU, Stream optimization, all of that plays a role in DS and BI seeing the same thing at the same time.

Basically, I got a higher percentage of motion rectangles lining up with BI after I optimized my streaming per the information in the Onboarding section of this forum. I hope this helps at some level...
  • "Whenever I take something apart to fix it and put it back together again, I end up with like six really important looking pieces left over" -Tim Allen
  • If you know what your after, you'll recognize it when you see it.
mike2246
Posts: 3
Joined: Wed Sep 15, 2021 2:05 pm

Re: Canceled Alert with 91% deepstack

Post by mike2246 »

I'm having similar problems with BI not sending a flag to DS to analyze the image/clip. I'm only having this problem with Reolink Cameras, my HIK and Amcrest work great using same Trigger/Motion/Record/Alert settings. In BI I see no alerts after walking out of the house to the car and back. Is there anything else that needs to be tweaked with a Reolink Camera?
Attachments
Settings 2.jpg
Settings 2.jpg (74.11 KiB) Viewed 5807 times
Settings 1.jpg
Settings 1.jpg (200.01 KiB) Viewed 5807 times
Fitz2380
Posts: 8
Joined: Thu Sep 24, 2020 10:59 pm

Re: Canceled Alert with 91% deepstack

Post by Fitz2380 »

mike2246 Did you ever get your Reolink cameras to successfully work. I too have the 410 series of cameras and have been frustrated in getting these to work consistently.
User avatar
YrbkMgr
Posts: 587
Joined: Sun Nov 24, 2019 12:56 am
Location: Chicagoland

Re: Canceled Alert with 91% deepstack

Post by YrbkMgr »

Fitz2380 wrote: Fri Nov 19, 2021 12:59 pm mike2246 Did you ever get your Reolink cameras to successfully work. I too have the 410 series of cameras and have been frustrated in getting these to work consistently.
In my experience it is not a detector issue as long as you are getting camera triggers/recordings when you should. Tweaking the detector or recording tab is for making sure you record the motion you've defined. Deepstack handles what to do after the camera is triggered. So to solve a problem with capturing events, focus on the Trigger and Recording tab (camera). To solve a problem with not getting alerted or false cancels, focus on DS.

The way I think about it is... divide the world into two camps. First camp is detecting motion that causes the camera to trigger and record the way you want. Second camp is to fine tune what happens once the camera is triggered - analyze objects and decide what happens next. Deepstack analysis creates an "IF THEN ELSE" condition for alerting. If DS recognizes an object, what should it do? In summary, is the problem Event Capture or Object Recognition - that's the first question to ask.

When you run a clip through DS using Testing and Tuning, you are making sure that DS correctly identifies objects in that clip. So, if an object in frame is a dog, and DS sees it as a cow... maybe you want to include cow in your "To Confirm" box in AI settings. It's just looking at a past event to see what objects it recognizes. It is testing the recognition capabilities only. Therefore, Analyze with Deepstack from the playback viewer is only useful in helping you define what goes in that confirm box (or cancel) in AI settings.

If the problem is that DS isn't triggering an Alert Action when it should, the first question to answer is: Does DS see the object at the right time? When you ctrl-dbl click on the clip with the DS status open, and highlight rectangles, do they line up with the object? In the post above (https://blueirissoftware.com/forum/view ... 526#p10526), I share what I had gleaned about how to review the DS status window. Since my post, BI Support have updated their Deepstack Fine Tuning page in the board index - worth looking at if you really want to troubleshoot DS problems.

As my best friend always says... don't listen to me, I'm not saying I'm right - I'm just saying how I see it.
  • "Whenever I take something apart to fix it and put it back together again, I end up with like six really important looking pieces left over" -Tim Allen
  • If you know what your after, you'll recognize it when you see it.
Post Reply