Closed blacknell closed 2 years ago
If the live view is "still" and not updating then that's what frigate sees too and would explain why things are being missed because the camera stream isn't working. Also keep in mind frigate does not create events on motion, only with true positive detected objects.
I also see you're using a reolink camera which is known to have many issues. Please refer to the docs on the config to attempt to improve this https://docs.frigate.video/configuration/camera_specific/#reolink-410520-possibly-others
Ok. Will try those settings again later (I did see this mentioned and tried something like this without success) and report back.
I should point out
I think something is related to my detect settings. I've told it (if I'e understood it right) to do the detection from the low res stream 640x360 and yet defined the resolution at 1280x720; and the recording at the hires stream which is 4k. Also, won't the default object sizes be a factor?
Ok. Will try those settings again later (I did see this mentioned and tried something like this without success) and report back.
Okay good to know, at the points where the web ui steam isn't updating then the ffmpeg process for that camera has errored out and is restarting.
I should point out
- the camera seems to be streaming well - if I put the same URL's into VNC they work fine.
To be clear VLC is a good tool and indicator but if it works in VLC isn't a 100% guarantee it should / will work with ffmpeg in frigate
- The video recordings it is creating do seem to be genuine ie they have a detected motion and also the box surrounding them categorising them as a person or car
Right, odds are when the camera works it works but the stream seems to be restarting which causes it to miss some things.
I think something is related to my detect settings. I've told it (if I'e understood it right) to do the detection from the low res stream 640x360 and yet defined the resolution at 1280x720; and the recording at the hires stream which is 4k. Also, won't the default object sizes be a factor?
Okay, you defined the resolution as different than the stream, ffmpeg will resize it manually to the detect size you specify. That will use more CPU but shouldn't affect object detection negatively. I see the higher quality stream recording, that is good. What do you mean by default object sizes?
What do you mean by default object sizes?
I thought I read in the notes something about defining the size of detectable objects relative to the image size, but now I can't find it.
What do you mean by default object sizes?
I thought I read in the notes something about defining the size of detectable objects relative to the image size, but now I can't find it.
You may be referring to this section: https://docs.frigate.video/guides/camera_setup which talks about the fact that the portion of the camera image is resized to 320x320 before object detection is done. The size of that is not configurable but yes the camera resolution affects it although just for smaller objects. Detecting cars or people that aren't on the other side of the street (just as an example) should be no issue.
Ok. I'm going to abandon the Reolink and buy a Dahua. In fact I already have, should be here later this week.
Regarding a Coral, the guidance points to any of the Coral products but can I run it all on the Dev Board Mini (on Mendel Linux) or must I use a co-processor type via USB or PCI?
@NickM-27 - ok, so I played around with Reolink settings and none of them worked with my Reolink 842A camera. Very disappointed (in the camera). FYI, seems to be an error in that camera setup above in that there are missing arguments - should be at least hwaccel_args: []
.
Would appreciate a tip above on the Coral board
That's unfortunate, that new camera should be quite a bit better.
As far as the coral goes, I believe it ~CAN~ but run on the dev board mini but I have seen a couple issues posted from users of them not working well. I am not knowledgeable in that regard so I am not sure exactly how viable that would be.
My Dahua camera arrived and things definitely improved. But the frame rates were definitely very poor and hence loads of events were being missed. Then it occurred to me the docker image I was running was wrong. I'm running it on my Mac with Apple Silicon and had specified the wrong architecture. When I took out the tag and just specified blakeblackshear/frigate:release-0.11.0-cbf26e0
things improved dramatically. Even detected at 4K was giving my 5fps and 12ms inference speed.
I would suggest running blakeblackshear/frigate:0.10.1-aarch64
instead. The 0.11.0 build is still experimental.
Describe the problem you are having
I've tried to keep the set up as simple as possible but I'm experiencing
I'd like to invest in a coral but a bit reluctant until I can prove to myself that it's working in principle
I'm happy to rerun whatever tests you want with debug flags etc.
Version
0.10.0-bfecee9
Frigate config file
Relevant log output
FFprobe output from your camera
Frigate stats
Operating system
Other
Install method
Docker Compose
Coral version
CPU (no coral)
Network connection
Wired
Camera make and model
Reolink 842A
Any other information that may be helpful
I'm running in docker compose with mqtt on the LAN. I have not connected this to HA yet. I also can't get the frigate home screen webcam image to stream (it's always a still shot)