Coral Edge USB not operating/recognised

[billynlazer]

So my new Coral Edge USB TPU has arrived, I changed the docker command to add the line -v /dev/bus/usb:/dev/bus/usb --privileged , eagerly hoping that it would be virtually plug and play.

My linux version shown by lsb_release -d gives result Ubuntu 22.04.3 LTS

But Ive got an ugly stream of repeating errors

[2023-08-08 19:08:46] [ERROR ] [viseron.components.ffmpeg.stream.camera_2] - [hevc @ 0x55fbb27a21c0] Could not find ref with POC 20
[2023-08-08 19:08:46] [ERROR ] [viseron.components.ffmpeg.stream.camera_2] - [hevc @ 0x55fbb27a21c0] Failed to end picture decode issue: 23 (internal decoding error).
[2023-08-08 19:08:46] [ERROR ] [viseron.components.ffmpeg.stream.camera_2] - [hevc @ 0x55fbb27a21c0] hardware accelerator failed to decode picture
[2023-08-08 19:08:46] [ERROR ] [viseron.components.ffmpeg.stream.camera_2] - Error while decoding stream #0:0: Input/output error

The first part of my config is as below

ffmpeg:
camera:
camera_2: # This value has to be unique across all cameras
name: Driveway
host: 192.168.1.80
port: 554
path: rtsp://192.168.1.80:554/Streaming/Channels/101?transportmode=unicast&profile=Profile_1
username: username
password: password
recorder:
retain: 2

Does anyone have any tips about how this might be addressed? Looks like the Coral is not being recognised and/or definitely not working :(

[billynlazer]

After posting this, I've switched over to another of my cameras, and that one seems to operate without errors, so I'm hoping that its a problem unique to the rtsp stream from the first camera. The second camera I tried didnt have the frame rate set to the same as the config in Viseron, which lead to some errors and a slow motion recording - syncing the frame rates seems to have fixed both those problems.

Still not entirely sure its utilizing the Coral - ffmpeg is sitting at around 20% cpu utilisation both with and without -v /dev/bus/usb:/dev/bus/usb --privileged , but then again, I dont know whether this is symptomatic of a not operating Coral - definitely not an expert here :)

[roflcoopter]

Your first error is not related to the Coral, your path is slightly wrong.
It should be path: /Streaming/Channels/101?transportmode=unicast&profile=Profile_1
Viseron will add together the host, port, path etc to form the full url.

FFmpeg does not use the Coral, it is only the edgetpu component that uses the Coral for object detection or image classification.

[billynlazer]

Ahhh. Thank you for your assistance!

I've made the correction and its stopped the infinite throwing of errors.

But it seems to be really insensitive in detecting a person - it will... but only when I'm 2-3 metres away from the camera, if I'm 7-8 metres away it misses me and doesnt trigger a recording. I've tried increasing the sensitivity, but it doesnt seem to solve that problem (and causes others).

At the moment for the config below, it triggers a record, but never actually stops recording, even long after I've left the scene

[2023-08-09 20:33:33] [INFO ] [viseron.components.nvr.nvr.camera_2] - Stopping recording in: 8
[2023-08-09 20:33:46] [INFO ] [viseron.components.nvr.nvr.camera_2] - Stopping recording in: 9, message repeated 2 times
[2023-08-09 20:33:47] [INFO ] [viseron.components.nvr.nvr.camera_2] - Stopping recording in: 8
[2023-08-09 20:34:14] [INFO ] [viseron.components.nvr.nvr.camera_2] - Stopping recording in: 9, message repeated 6 times
[2023-08-09 20:34:15] [INFO ] [viseron.components.nvr.nvr.camera_2] - Stopping recording in: 8
[2023-08-09 20:34:36] [INFO ] [viseron.components.nvr.nvr.camera_2] - Stopping recording in: 9, message repeated 5 times
[2023-08-09 20:34:37] [INFO ] [viseron.components.nvr.nvr.camera_2] - Stopping recording in: 8
[2023-08-09 20:34:38] [INFO ] [viseron.components.nvr.nvr.camera_2] - Stopping recording in: 7
[2023-08-09 20:34:39] [INFO ] [viseron.components.nvr.nvr.camera_2] - Stopping recording in: 6
[2023-08-09 20:35:00] [INFO ] [viseron.components.nvr.nvr.camera_2] - Stopping recording in: 9, message repeated 3 times
[2023-08-09 20:35:01] [INFO ] [viseron.components.nvr.nvr.camera_2] - Stopping recording in: 8
[2023-08-09 20:35:03] [INFO ] [viseron.components.nvr.nvr.camera_2] - Stopping recording in: 7

config is...

ffmpeg:
camera:

camera_2:  # This value has to be unique across all cameras
  name: Backyard
  host: 192.168.1.64
  port: 554
  path: /Streaming/Channels/101?transportmode=unicast&profile=Profile_1
  username: <usr>
  password: <pwd>
  fps: 8
  recorder:
    retain: 2
    codec: h264

darknet:
object_detector:
cameras:

  camera_2:  # Attach detector to the configured camera_2 above
    fps: 2
    scan_on_motion_only: true
    labels:
      - label: person
        confidence: 0.65
        trigger_recorder: true

mog2:
motion_detector:
cameras:
camera_2: # Attach detector to the configured camera_2 above
trigger_recorder: false
threshold: 10
fps: 2
area: 0.04

nvr:
camera_2: # Run NVR for camera_2

Is there something wrong with this config? the idea is that when the motion detector detects movement, it doesnt turn on recording, but instead the object detector turns on. IF the object detector records a person, then an actual recording is made. At least thats the plan...

[roflcoopter]

Sorry i have been away on vacation.

If you enable debug logging for darknet and mog2 it will be a bit easier to figure out what is going on.
Your config is correct for what you want to do, have motion detection trigger object detection, which then will trigger recordings if there is an object of interest in the image.

You can add debug logging to specific components like this:

logger:
  logs:
    viseron.components.mog2: debug
    viseron.components.darknet: debug
    viseron.components.nvr: debug

[billynlazer]

No rush to reply, I appreciate your help.

Here is the log of a couple of attempts - from inspection, it looks as if walking in front of the camera actually causes a termination, then docker restarts the container.

No recording is every written to disk... maybe it gets written to memory and then exits before writing out to disk.

I'm running the docker container as root, so priviledges shouldnt be a problem.

log2.log

[roflcoopter]

I just noticed that you are using the darknet component for object detection.
darknet does not use the Coral.

Since the container seems to restart, i am thinking it could be a memory issue.
What are the specs of your host machine?