[Detector Support]: ONNX YoloV9 with iGPU but low usage #20449

LucaCraft89 · 2025-10-12T15:32:17Z

LucaCraft89
Oct 12, 2025

Describe the problem you are having

I have low gpu usage when using ONNX with YoloV9 size M while CPU takes most of the work, why?

also i can't get face and licence plate recognition to work

Version

0.16.1-e664cb2

Frigate config file

############
# Frigate Configuration File
############

mqtt:
  host: 192.168.1.256
  port: 1883
  user: FDAS)DIPADSOJ
  password: IOUHADOIHASDOIHASD
  topic_prefix: frigate

detectors:
  # ov:
  #   type: openvino
  #   device: GPU
  onnx:
    type: onnx
    device: GPU

model:
  model_type: yolo-generic
  width: 320 # <--- should match the imgsize set during model export
  height: 320 # <--- should match the imgsize set during model export
  input_tensor: nchw
  input_dtype: float
  path: /config/model_cache/yolov9-m-320.onnx
  labelmap_path: /labelmap/coco-80.txt
# model:
#   width: 300
#   height: 300
#   input_tensor: nhwc
#   input_pixel_format: bgr
#   path: /openvino-model/ssdlite_mobilenet_v2.xml
#   labelmap_path: /openvino-model/coco_91cl_bkgr.txt

#model: # required
#  # name of model (will be automatically downloaded) or path to your own .rknn model file
#  # possible values are:
#  # - frigate-fp16-yolov9-t
#  # - frigate-fp16-yolov9-s
#  # - frigate-fp16-yolov9-m
#  # - frigate-fp16-yolov9-c
#  # - frigate-fp16-yolov9-e
#  # your yolo_model.rknn
#  path: frigate-fp16-yolov9-t
#  model_type: yolo-generic
#  width: 320
#  height: 320
#  input_tensor: nhwc
#  labelmap_path: /labelmap/coco-80.txt

go2rtc:
  streams:
    cam1:
      - rtsp://192.168.1.184/onvif1
      - ffmpeg:cam1#audio=opus
    cam2:
      - rtsp://192.168.1.182/user=admin&password=tlJwpbo6&channel=1&stream=0.sdp?real_stream
      - ffmpeg:cam2#audio=opus
    cam3:
      - rtsp://192.168.1.181/user=admin&password=tlJwpbo6&channel=1&stream=0.sdp?real_stream
      - ffmpeg:cam3#audio=opus
    cam4:
      - rtsp://192.168.1.183/user=admin&password=tlJwpbo6&channel=1&stream=0.sdp?real_stream
      - ffmpeg:cam4#audio=opus

objects:
  track:
    - person
    - car
    - truck
    - motorcycle
    - bicycle
    - dog
    - cat
    - bird
    - face
    - license_plate
    - package
  #filters:
    #person:
  #    min_area: 5000
   #   max_area: 100000
  #    min_score: 0.5
   #   threshold: 0.7
 #   car:
 #     min_area: 5000
  #   max_area: 200000
  #    min_score: 0.5
  #    threshold: 0.7

record:
  enabled: true
  retain:
    days: 7     # Number of days to keep recordings
    mode: all   # 'all' to save everything or 'events' to only save clips with detected objects

snapshots:
  enabled: true
  clean_copy: true   # Save a clean copy without bounding boxes
  timestamp: true    # Add timestamp to snapshots
  bounding_box: true # Draw bounding boxes on objects
  retain:
    default: 14      # Number of days to keep snapshots

birdseye:
  enabled: true
  mode: continuous   # Options: continuous, motion, or objects
  width: 1280
  height: 720

# Camera configurations
cameras:
  cam1:  # Entry Camera
    ffmpeg:
      output_args:
        record: preset-record-generic-audio-aac
      inputs:
        - path: rtsp://127.0.0.1:8554/cam1
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
            - audio
      path: '5.0'
    detect:
      enabled: true
   #   width: 640  # Width for detection (lower for better performance)
   #   height: 360 # Height for detection
   #   fps: 5      # Frames per second used for detection
    # objects:
    #   track:
    #     - person
    #     - car
    #     - truck
    #     - motorcycle
    #     - bicycle
    #     - dog
    #     - cat
    #     - bird

    #  filters:
     #   person:
  #        min_area: 5000
    #      max_area: 100000
  #        min_score: 0.5
  #        threshold: 0.7
    record:
      enabled: true
      retain:
        days: 7
        mode: all
    snapshots:
      enabled: true
      retain:
        default: 14

    motion:
      threshold: 39
      contour_area: 17
      improve_contrast: true
      mask: 0,0.062,0.392,0.062,0.394,0,0,0
    zones:
      Garage:
        coordinates: 0.483,0.421,0.003,0.473,0,1,1,1,0.993,0.452
        loitering_time: 0
    review:
      alerts:
        required_zones: Garage
  cam2:  # Side Camera
    ffmpeg:
      output_args:
        record: preset-record-generic-audio-aac
      inputs:
        - path: rtsp://127.0.0.1:8554/cam2
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
            - audio
    onvif:
      host: 192.168.1.182
      port: 8899
      user: admin
      password: tlJwpbo6

    detect:
      enabled: true
    #  width: 960
    #  height: 1080  # Adjusted for stacked 16:9 streams
    #  fps: 5
    # objects:
    #   track:
    #     - person
    #     - car
    #     - truck
    #     - motorcycle
    #     - bicycle
    #     - dog
    #     - cat
    #     - bird
    #  filters:
    #    person:
    #      min_area: 5000
    #      max_area: 100000
    #      min_score: 0.5
    #      threshold: 0.7
    record:
      enabled: true
      retain:
        days: 7
        mode: all
    snapshots:
      enabled: true
      retain:
        default: 14

    motion:
      threshold: 40
      contour_area: 33
      improve_contrast: true
    zones:
      Front_Yard:
        coordinates: 
          0.673,0,0.662,0.207,0.652,0.252,0.533,0.229,0.543,0.305,0.492,0.374,0.486,0.408,0.437,0.44,0.435,0.501,0.999,0.501,1,0.18,0.898,0
        loitering_time: 0
        inertia: 3
      Side_Yard:
        coordinates: 
          0,0.499,0.429,0.499,0.438,0.432,0.481,0.411,0.489,0.371,0.545,0.308,0.528,0.225,0.647,0.249,0.658,0.206,0.67,0.001,0.001,0.003
        loitering_time: 0
    review:
      alerts:
        required_zones:
          - Front_Yard
          - Side_Yard
  cam3:  # Backyard Camera
    ffmpeg:
      output_args:
        record: preset-record-generic-audio-aac
      inputs:
        - path: rtsp://127.0.0.1:8554/cam3
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
            - audio
    onvif:
      host: 192.168.1.181
      port: 8899
      user: admin
      password: tlJwpbo6
    detect:
      enabled: true
    #  width: 960
    #  height: 1080  # Adjusted for stacked 16:9 streams
    #  fps: 5
    # objects:
    #   track:
    #     - person
    #     - car
    #     - truck
    #     - motorcycle
    #     - bicycle
    #     - dog
    #     - cat
    #     - bird
    #  filters:
    #    person:
  #        min_area: 5000
    #      max_area: 100000
   #       min_score: 0.5
   #       threshold: 0.7
    record:
      enabled: true
      retain:
        days: 7
        mode: all
    snapshots:
      enabled: true
      retain:
        default: 14

    zones:
      Back_Yard:
        coordinates: 
          0.406,0.496,0.001,0.497,-0.001,0.316,0.048,0.268,0.263,0.255,0.274,0.079,0.548,0.034,0.525,0.499
        loitering_time: 0
    review:
      alerts:
        required_zones: Back_Yard
  cam4:  # Front Camera
    ffmpeg:
      output_args:
        record: preset-record-generic-audio-aac
      inputs:
        - path: rtsp://127.0.0.1:8554/cam4
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
            - audio
    onvif:
      host: 192.168.1.183
      port: 8899
      user: admin
      password: tlJwpbo6
    detect:
      enabled: true
    #  width: 960
    #  height: 1080  # Adjusted for stacked 16:9 streams
    #  fps: 5
    objects:
      # track:
      #   - person
      #   - car
      #   - truck
      #   - motorcycle
      #   - dog
      #   - cat
      #   - bird
      filters:
    #    person:
  #       min_area: 5000
  #        max_area: 100000
  #        min_score: 0.5
  #        threshold: 0.7
        bicycle:
          mask:
            - 0.588,0.339,0.818,0.347,0.83,0.245,0.772,0.22,0.705,0.215,0.7,0.187,0.649,0.186,0.63,0.225,0.585,0.241
            - 0.287,0.665,0.558,0.662,0.532,0.588,0.496,0.569,0.456,0.537,0.418,0.533,0.365,0.518,0.335,0.561,0.29,0.562
        car:
          mask: 
            0.003,0.446,0.117,0.365,0.292,0.207,0.51,0.143,0.718,0.161,0.785,0.274,0.758,0.434,0.733,0.47,0.71,0.474,0.699,0.499,0.001,0.499
    record:
      enabled: true
      retain:
        days: 7
        mode: all
    snapshots:
      enabled: true
      retain:
        default: 14

    zones:
      Exterior_Car_Spot:
        coordinates: 
          0.38,0.496,0.816,0.501,0.864,0.32,0.893,0.163,0.525,0.14,0.527,0.31
        inertia: 3
        loitering_time: 0
        objects:
          - bird
          - cat
          - dog
          - motorcycle
          - person
          # - truck
      Gates:
        coordinates: -0.001,0.501,0.38,0.497,0.525,0.309,0.527,0.142,0,0.146
        inertia: 3
        loitering_time: 0
    review:
      alerts:
        required_zones:
          - Exterior_Car_Spot
          - Gates
ffmpeg:
  # hwaccel_args: "auto"
  input_args: preset-rtsp-generic
  output_args:
    record: preset-record-generic-audio-aac
  # path: "5.0"
  hwaccel_args: auto
  # global_args:
  #   - -hide_banner
  #   - -loglevel
  #   - warning
  # input_args:
  #   - -avoid_negative_ts
  #   - make_zero
  #   - -fflags
  #   - nobuffer+genpts+discardcorrupt
  #   - -flags
  #   - low_delay
  #   - -strict
  #   - experimental
  #   - -analyzeduration
  #   - '1000000'
  #   - -probesize
  #   - '1000000'
  # output_args:
  #   record:
  #     - -f
  #     - segment
  #     - -segment_time
  #     - '60'
  #     - -segment_format
  #     - mp4
  #     - -reset_timestamps
  #     - '1'
  #     - -strftime
  #     - '1'
  #     - -c
  #     - copy

# Increase shm-size as recommended in logs
# This must be set in your docker-compose.yml or docker run command
# docker run --shm-size=98m ...

# Optional semantic search feature
semantic_search:
  enabled: true
  reindex: false
  model_size: large
version: 0.16-0
notifications:
  enabled: true
  email: [email protected]
detect:
  enabled: true
face_recognition:
  enabled: true
  model_size: large
lpr:
  enabled: true
classification:
  bird:
    enabled: true

docker-compose file or Docker CLI command

version: "3.9"
services:
  frigate:
    container_name: frigate
    restart: unless-stopped
    stop_grace_period: 30s
    image: ghcr.io/blakeblackshear/frigate:stable
    shm_size: '2048m'
    devices:
      - /dev/dri/renderD128:/dev/dri/renderD128
    volumes:
      - ./config:/config
      - ./storage:/media/frigate
      - type: tmpfs # Optional: 1GB of memory, reduces SSD/SD Card wear
        target: /tmp/cache
        tmpfs:
          size: 3000000000
    cap_add:
      - CAP_PERFMON
    ports:
      - "8971:8971"
      - "8554:8554" # RTSP feeds
      - "5000:5000"

Relevant Frigate log output

2025-10-12 15:23:22.872242305  [INFO] Preparing Frigate...
2025-10-12 15:23:23.683556209  [INFO] Starting Frigate...
2025-10-12 15:23:26.431867631  [2025-10-12 15:23:26] frigate.util.config            INFO    : Checking if frigate config needs migration...
2025-10-12 15:23:26.479474241  [2025-10-12 15:23:26] frigate.util.config            INFO    : frigate config does not need migration...
2025-10-12 15:23:26.853157329  [2025-10-12 15:23:26] frigate.util.services          INFO    : Automatically detected vaapi hwaccel for video decoding
2025-10-12 15:23:35.731357729  [2025-10-12 15:23:35] frigate.config.config          WARNING : cam1 is configured to track ['package', 'face', 'truck', 'license_plate'] objects, which are not supported by the current model.
2025-10-12 15:23:37.672931014  [2025-10-12 15:23:37] frigate.config.config          WARNING : cam2 is configured to track ['package', 'face', 'truck', 'license_plate'] objects, which are not supported by the current model.
2025-10-12 15:23:39.637420425  [2025-10-12 15:23:39] frigate.config.config          WARNING : cam3 is configured to track ['package', 'face', 'truck', 'license_plate'] objects, which are not supported by the current model.
2025-10-12 15:23:41.592182520  [2025-10-12 15:23:41] frigate.config.config          WARNING : cam4 is configured to track ['package', 'face', 'truck', 'license_plate'] objects, which are not supported by the current model.
2025-10-12 15:23:41.596785680  [2025-10-12 15:23:41] frigate.app                    INFO    : Starting Frigate (0.16.1-e664cb2)
2025-10-12 15:23:41.617337221  [2025-10-12 15:23:41] peewee_migrate.logs            INFO    : Starting migrations
2025-10-12 15:23:41.641488549  [2025-10-12 15:23:41] peewee_migrate.logs            INFO    : There is nothing to migrate
2025-10-12 15:23:41.665616430  [2025-10-12 15:23:41] frigate.app                    INFO    : Recording process started: 557
2025-10-12 15:23:41.677829651  [2025-10-12 15:23:41] frigate.app                    INFO    : Review process started: 570
2025-10-12 15:23:41.685291139  [2025-10-12 15:23:41] frigate.app                    INFO    : go2rtc process pid: 123
2025-10-12 15:23:41.710529230  [2025-10-12 15:23:41] detector.onnx                  INFO    : Starting detection process: 582
2025-10-12 15:23:41.718961260  [2025-10-12 15:23:41] frigate.detectors.plugins.onnx INFO    : ONNX: loaded onnxruntime module
2025-10-12 15:23:41.720752616  [2025-10-12 15:23:41] frigate.detectors.plugins.onnx INFO    : ONNX: loading /config/model_cache/yolov9-m-320.onnx
2025-10-12 15:23:41.721810754  [2025-10-12 15:23:41] frigate.app                    INFO    : Embedding process started: 585
2025-10-12 15:23:41.755124362  [2025-10-12 15:23:41] frigate.app                    INFO    : Output process started: 612
2025-10-12 15:23:41.811097870  [2025-10-12 15:23:41] frigate.app                    INFO    : Camera processor started for cam1: 633
2025-10-12 15:23:41.838033784  [2025-10-12 15:23:41] frigate.app                    INFO    : Camera processor started for cam2: 640
2025-10-12 15:23:42.044427714  [2025-10-12 15:23:41] frigate.app                    INFO    : Camera processor started for cam3: 646
2025-10-12 15:23:42.052559559  [2025-10-12 15:23:42] frigate.app                    INFO    : Camera processor started for cam4: 649
2025-10-12 15:23:42.102673257  [2025-10-12 15:23:42] frigate.app                    INFO    : Capture process started for cam1: 716
2025-10-12 15:23:42.122511369  [2025-10-12 15:23:42] frigate.app                    INFO    : Capture process started for cam2: 726
2025-10-12 15:23:42.138066743  [2025-10-12 15:23:42] frigate.app                    INFO    : Capture process started for cam3: 736
2025-10-12 15:23:42.177765889  [2025-10-12 15:23:42] frigate.app                    INFO    : Capture process started for cam4: 742
2025-10-12 15:23:42.236845348  [2025-10-12 15:23:42] frigate.detectors.plugins.onnx INFO    : ONNX: /config/model_cache/yolov9-m-320.onnx loaded
2025-10-12 15:23:42.343996873  [2025-10-12 15:23:42] frigate.api.fastapi_app        INFO    : Starting FastAPI app
2025-10-12 15:23:42.541001927  [2025-10-12 15:23:42] frigate.api.fastapi_app        INFO    : FastAPI started
2025-10-12 15:23:44.026876126  loading data from : /config/model_cache/facedet/landmarkdet.yaml
2025-10-12 15:23:46.219642415  INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
2025-10-12 15:24:23.179591559  [2025-10-12 15:24:23] frigate.api.media              ERROR   : No recordings found for cam2 during the requested time range
2025-10-12 15:24:23.179591559  [LOGGING] Last message repeated 2 times
2025-10-12 15:28:06.267909582  [2025-10-12 15:28:06] frigate.ptz.onvif              INFO    : Attempting ONVIF initialization for cam2 (retry 1/5)
2025-10-12 15:28:06.842682697  [2025-10-12 15:28:06] frigate.ptz.onvif              INFO    : Attempting ONVIF initialization for cam4 (retry 1/5)
2025-10-12 15:28:07.098921159  [2025-10-12 15:28:07] frigate.ptz.onvif              INFO    : Attempting ONVIF initialization for cam3 (retry 1/5)

Install method

Docker Compose

Object Detector

OpenVino

Screenshots of the Frigate UI's System metrics pages

Any other information that may be helpful

this is a laptop with 8th gen i7 and iGPU

Answered by NickM-27

Oct 12, 2025

Your goal shouldn't be to max out GPU usage. If the model runs well and meets the detection needs then it's best to let the GPU run for other features.

View full answer

2025-10-12T15:32:43Z

inkeep-ai-assistant[bot]
bot Oct 12, 2025

Based on your configuration and the logs, there are several issues to address:

GPU Usage Issue

Your ONNX detector is configured with device: GPU, but for Intel iGPU, you should use the OpenVINO detector instead. The recommended detector type for Intel hardware is openvino with yolov9 models(1).

Try this configuration:

detectors:
  ov:
    type: openvino
    device: GPU

model:
  model_type: yolov9
  width: 320
  height: 320
  input_tensor: nchw
  input_dtype: float
  path: /config/model_cache/yolov9-m-320.onnx

(2)

Face and License Plate Recognition Issues

The warnings in your logs indicate the problem:

cam1 is configured to track ['package', 'face', 'truck', 'license_plate'] objects, which are not supported by the current model.

Your YOLOv9 model uses the standard COCO-80 labelmap, which doesn't include face, license_plate, package, or truck(2).

For face and license plate recognition to work with your current setup:

Remove unsupported objects from your tracking list:

objects:
  track:
    - person
    - car
    - motorcycle
    - bicycle
    - dog
    - cat
    - bird

(3)

License Plate Recognition will work automatically when cars are detected and moving. LPR uses a separate YOLOv9 license plate detection model that runs when car or motorcycle objects are tracked(3).
Face Recognition requires detected person objects to work(3).

The current model limitations mean you won't see face, license_plate, package, or truck as tracked objects, but the enrichment features (LPR and face recognition) will still function when their prerequisite objects (car/motorcycle for LPR, person for face recognition) are detected.

📚 Sources:

Have another question? Just tag @inkeep.

0 replies

LucaCraft89 · 2025-10-12T15:36:38Z

LucaCraft89
Oct 12, 2025
Author

Changed to

detectors:
  ov:
    type: openvino
    device: GPU
  # onnx:
  #   type: onnx
  #   device: GPU

model:
  model_type: yolo-generic
  width: 320 # <--- should match the imgsize set during model export
  height: 320 # <--- should match the imgsize set during model export
  input_tensor: nchw
  input_dtype: float
  path: /config/model_cache/yolov9-m-320.onnx
  labelmap_path: /labelmap/coco-80.txt

now i get more consistant usage

i belive there is still more to give, bigger model?

13 replies

NickM-27 Oct 12, 2025
Collaborator Sponsor

Your goal shouldn't be to max out GPU usage. If the model runs well and meets the detection needs then it's best to let the GPU run for other features.

Answer selected by LucaCraft89

NickM-27 Oct 12, 2025
Collaborator Sponsor

In this case your inference speed is also higher than recommended already, there may be issues tracking objects reliably

LucaCraft89 Oct 12, 2025
Author

what do you mean with that?
also do i need to do extra stuff to get face and licence plate recognition to work?

NickM-27 Oct 12, 2025
Collaborator Sponsor

Which part are you asking about? I was referring to the inference speed

And for face rec you need to create at least one face library in the UI

Currently LPR doesn't run on stationary objects

LucaCraft89 Oct 12, 2025
Author

sorry maybe i did not explain myself,

what is inference speed, do i need to worry about it?

i can't get face recognition to work even after following the instructions, any other steps?

LPR never seen it work but i do not mind it

2025-10-12T16:50:15Z

inkeep-ai-assistant[bot]
bot Oct 12, 2025

Inference speed is the time (in milliseconds) a detector takes to process a frame(1). This is a key metric for sizing hardware and balancing accuracy versus latency(2).

From your screenshot showing "Velocità inferenza rilevatore" (inference speed detector) at 65.0ms, this is higher than ideal. Higher inference speeds can cause issues with object tracking reliability, as the system may struggle to keep up with real-time processing.

For optimal performance, you want lower inference speeds. The current speed suggests your model may be too heavy for your hardware, or you might benefit from:

Using a smaller model size
Reducing the detect resolution
Lowering the detect FPS

Your GPU usage looks good now with the OpenVINO detector, but the focus should be on stable, reliable detection rather than maximizing GPU utilization.

📚 Sources:

Have another question? Just tag @inkeep.

0 replies

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

[Detector Support]: ONNX YoloV9 with iGPU but low usage #20449

Uh oh!

{{title}}

Uh oh!

Replies: 3 comments 13 replies

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{editor}}'s edit

{{editor}}'s edit

Uh oh!

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{title}}

Uh oh!

Select a reply

Uh oh!

Uh oh!

[Detector Support]: ONNX YoloV9 with iGPU but low usage #20449

Uh oh!

LucaCraft89 Oct 12, 2025

Describe the problem you are having

Version

Frigate config file

docker-compose file or Docker CLI command

Relevant Frigate log output

Install method

Object Detector

Screenshots of the Frigate UI's System metrics pages

Any other information that may be helpful

Replies: 3 comments · 13 replies

Uh oh!

inkeep-ai-assistant[bot] bot Oct 12, 2025

GPU Usage Issue

Face and License Plate Recognition Issues

Uh oh!

LucaCraft89 Oct 12, 2025 Author

Uh oh!

NickM-27 Oct 12, 2025 Collaborator Sponsor

Uh oh!

Uh oh!

NickM-27 Oct 12, 2025 Collaborator Sponsor

Uh oh!

LucaCraft89 Oct 12, 2025 Author

Uh oh!

NickM-27 Oct 12, 2025 Collaborator Sponsor

Uh oh!

LucaCraft89 Oct 12, 2025 Author

Uh oh!

inkeep-ai-assistant[bot] bot Oct 12, 2025

LucaCraft89
Oct 12, 2025

Replies: 3 comments 13 replies

inkeep-ai-assistant[bot]
bot Oct 12, 2025

LucaCraft89
Oct 12, 2025
Author

NickM-27 Oct 12, 2025
Collaborator Sponsor

NickM-27 Oct 12, 2025
Collaborator Sponsor

LucaCraft89 Oct 12, 2025
Author

NickM-27 Oct 12, 2025
Collaborator Sponsor

LucaCraft89 Oct 12, 2025
Author

inkeep-ai-assistant[bot]
bot Oct 12, 2025