I don’t really understand the gasket driver part, is this difficult to do ? and is this something performed in hass.io ?
That’s cool, looks like they basically doubled the original circuit. I wonder if bifurcation will be required for that. They do state that it might not be compatible because of the two lanes. Also they recommend adding a heatsink.
I might pick one up, my mobo is an E key. Although right now I have four cameras and it’s not breaking a sweat.
I don’t have frigate running in hass.io, so I’m not sure if a DKMS driver can be installed for that. My guess would be that it is possible, but could be tricky.
Last time I looked, hassos didnt even include the kernel modules to enable use of intel quicksync to accelerate decoding in ffmpeg. I don’t imagine the gasket drive is in there either. If you are running hassio on a different host os, then you should be able to install it according to the instructions.
Hi, I am trying to get frigate up and running with the coral but am running into the same error messages:
Detection appears to have stopped. Restarting detection process
Starting detection process: 45
…etc…
Did you ever resolve your issue?
When I run dmesg I see a string of these errors:
traps: python3.7[10785] trap invalid opcode ip:7fe80c4c9be4 sp:7fe806e4ea50 error:0 in libedgetpu.so.1.0[7fe80c47c000+e2000]
Does anybody know what this means? Thanks!
My understanding is that the NUC image of supervised has had the ability for hardware acceleration of FFMPEG added to it. If not, you could add it on OS level. But yes, with hassos it will be difficult.
Just posted RC2 for 0.6.0.
- Updated method for computing a median score that reduces false positives
- New min_score and threshold definitions
- Allows configuring a specific EdgeTPU device for the container
- Modified labelmap to treat trucks as car to avoid confusion between the two
- Crop and resize the best.jpg endpoint
- Crop and resize images before sending to MQTT
- Limit what object types trigger recordings
- Limit cache size
- Polygon mask type
- Added affiliate links to the readme
I sorted through nearly 3k clips to identify false and true positives and landed on a methodology that filtered out 95% of the false positives that occurred over a 3 week period across all my cameras. It only resulted in a loss of 8% of the true positives. Here were my results at different thresholds.
8 Likes
Some great stuff in 0.115 that will go a long way to using hass as an NVR replacement!
1 Like
I haven’t tested it myself, but I think you can run with the USB version of the coral. You might not get hw accelerated decoding of your video feeds. I am assuming you are on a NUC.
Yes! The media features in 0.115 look great!
I’m on a NUC, but in australia it’s impossible to find the USB version locally
Which version of opencv did you end up using? I’m having the same problem with another component.
cat Dockerfile|grep headless
opencv-python-headless==4.2.0.34
and i beg you to use Ubuntu 18.04 as base platform, no fancy stuff, this is rock solid, clean and also modern.
1 Like
NVR replacement would be a dream, but even if you are the best coder of the world we need to wait for nvidia and amd. I have just asked my nvidia contact today about the jetson hw accel support timeline. lets hope the best.
Blakeblackshear indicated it may be insufficient power to drive the Coral. I even bought a a PCI USB 3.0 card but was still receiving the same restart messages and have since given up.
Hi Blake, I have been running for a few days and then suddenly my camera_fps has increased from around 10fps to 100fps…camera looks ok on client app, any idea what might be happening. If I kill the container and the restart sometimes i goes straight back to 100fps and sometimes 10fps…currently running ok again now…strange?
If you are running 0.6.0, you may need to set FPS in the config for that camera.
i am i think, where is it in the config?
Actually I am not so sure now, I have take_frame in my config.yaml. I downloaded the frigate_master from github a few weeks ago…