Motion detection identification missed

Post Reply
Jey
Posts: 6
Joined: Fri Jul 30, 2021 4:18 pm

Motion detection identification missed

Post by Jey » Sat Aug 21, 2021 2:29 pm

Hi! :)

I installed BI (full licence) 3 weeks ago, with 2 Annke C800 (new model).
The main stream is 4k 12 fps, and the substream 640 x 480.

Since 3 weeks, I'm looking how it's working, and I change settings to improve.
It's working pretty well. The motion detection detect squirrel, rabbit and few things (but unfortunately, it's not identified by DeepStack AI). I have bear, lynx, wolf, moose, so I hope to catch 1 one day... :)

But it is not working well with car arriving on the parking or moving on the road (track).

For example, when I arrived by car (a big one, VW Caravelle) then walked to the house, DeepStack didn't identified anything. But when I came back and left it was ok (person + car).

And here it's the same problem with the postman :

He arrived by car and walked, the motion detection was ok but nothing was identified by DS AI (I just saw it by checking cancelled motion detections).
https://u.pcloud.link/publink/show?code ... iqU8feON07

Here, the postman came back and it was well identified (person 59% + car 85%)
https://u.pcloud.link/publink/show?code ... N64FOuPD6y

And here, this car moving slowly near on a track was never identified (or all other time that I didn't put in the video):
https://u.pcloud.link/publink/show?code ... lIm4kPhzYy

Yes this is a mess on the parking, but we have building works :D

This is my configuration:
Image
Image
Image
Image


Here each picture taken when motion was detected.
Image
Image
Image
Image


Is it possible to improve this detection ?
To "teach" to DeepStack that this was a car ?

Thanks for your help :)
User avatar
YrbkMgr
Posts: 129
Joined: Sun Nov 24, 2019 12:56 am

Re: Motion detection identification missed

Post by YrbkMgr » Mon Aug 23, 2021 11:23 pm

I'm trying to answer the same basic questions as you are and have been pouring over information ever since I installed DS a month ago. What I can tell you is that this is a common issue from what I've seen in the forums and is not straightforward in exactly how to troubleshoot it. That said, the following may give you some insight as to what's happening.

First, In order to analyze it, you have to first turn on Save DeepStack Analysis in the AI settings of the Trigger Tab. This causes DS to create a record of what it used for object detection as a DAT file, that is placed in your Alerts folder. This gives you a way to review what happened. Without that, you can't really troubleshoot.

If you right click on the playback and choose Testing and Tuning, then enable Analyze with DS, that will RE-analyze the entire field of view of the BVR recording - that is the same concept as Analyze with Motion Detector, which is used to fine tune your motion detector settings; it's temporary for fine tuning, not something that's on all the time. But that won't tell you what happened when the clip was created. I'm still grappling with using DS test n tune. Just know that it's different and used differently.

Once you have Save enabled for DS, then you open the Status Window of BI from the main console and select the DS tab. When you double click on the Alert from the Clips list on the left, the DAT file will populate data into the Status window and show the analysis used. This tells you, in theory, where DS analyzed (yellow boxes in the image) for objects, when it is analyzing that area (timestamp), any object it identified at the time, and finally, the image that DS used to confirm or cancel an alert - this image is denoted by the asterisk in the object tree on the left. Side note: I have no idea how to interpret the timestamps...yet.

When there is a mismatch between the yellow rectangle and the object that triggered the camera - areas don't overlap, an alert is cancelled. Usually.

While trying to figure this out for myself, I created a video that I posted on YT here: DeepStack Object Detection via Motion Trigger. It is just a quick visual of what I am trying to explain above. But the takeaway is...BI drives what area will be used by DS (yellow rectangles) for object identification.

Page 70 of the help file says: "You may also benefit from configuring the motion sensor to use more sensitive settings so that more events will be evaluated by the AI. This is done by lowering the object size and contrast, as well as possibly reverting to the “simple” detection algorithm."

So in summary, first turn on the Save dat part. Then review your cancelled alerts to see what area is being analyzed for motion and what DS identifies. The tricky part is figuring out what to change and where.

There's more to it, but that is how to start trying to figure out what's going on.

Disclaimer: The above are my conclusions from my own research and as such, I could be wrong about some things. There is a notion that the best way to learn about something is to try and explain it to someone else - that's pretty much what's going on here.
What I love about standards is that there are so many to choose from!
Jey
Posts: 6
Joined: Fri Jul 30, 2021 4:18 pm

Re: Motion detection identification missed

Post by Jey » Tue Aug 24, 2021 9:52 am

Thanks a lot for this long answer.
I will follow it soon when I will have time to work on BI.


I don't receive notification (when someone replies) by email (even if notification is enabled), and I don't have antispam on my email.
Is there something I missed?
User avatar
YrbkMgr
Posts: 129
Joined: Sun Nov 24, 2019 12:56 am

Re: Motion detection identification missed

Post by YrbkMgr » Tue Aug 24, 2021 4:52 pm

Jey wrote: Tue Aug 24, 2021 9:52 am Thanks a lot for this long answer.
I will follow it soon when I will have time to work on BI.


I don't receive notification (when someone replies) by email (even if notification is enabled), and I don't have antispam on my email.
Is there something I missed?
It's not you.... that feature is apparently broken in this forum. I don't receive notifications either and even sent a note to the forum Admin - no reply.

So you just have to check manually for the time being...
What I love about standards is that there are so many to choose from!
Post Reply