I have set up a custom model, but Frigate cannot recognize people

[13516513760]

My yml file content is:

mqtt:
  enabled: true
  port: 1883
  topic_prefix: frigate
  client_id: emqx_MTc1Nj
  host: host.docker.internal
  user: admin
  password: znyj@123456

auth:
  enabled: true

cameras:
  name_of_your_camera:
    enabled: true
    ffmpeg:
      inputs:
        - path: rtsp://admin:[email protected]:554/h264/ch1/main/av_stream
          roles:
            - detect
    detect:
      enabled: true
      fps: 1
      # ✅ 移除width/height，使用全局的320x320
    motion: {}
    snapshots:
      enabled: true
      # ✅ 临时移除required_zones，让所有检测都能保存
      # required_zones:
      #   - person_test
      #   - user_test
    record:
      enabled: true
      retain:
        days: 7

    # ✅ 临时移除zones，稍后再添加
    # zones: {}
    review:
      detections:
        labels:
          - person

    mqtt:
      enabled: true
      timestamp: true
      bounding_box: true

    # ✅ 摄像头级别只追踪person
    objects:
      track:
        - person
      filters:
        person:
          min_area: 100      # ✅ 极低，便于调试
          max_area: 999999
          threshold: 0.1     # ✅ 极低，便于调试

    zones:
      person_area:
        coordinates: 
          0.315,0.466,0.791,0.398,0.985,0.539,0.983,0.85,0.079,0.948,0.072,0.548
        loitering_time: 0
        objects: person
version: 0.16-0


detect:
  width: 320
  height: 320
  fps: 1
  stationary:
    interval: 1
    threshold: 25
  enabled: true

snapshots:
  enabled: true
  bounding_box: true
  timestamp: true
  retain:
    default: 10

semantic_search:
  enabled: false
  reindex: false
  model_size: small

detectors:
  onnx_cpu:
    type: onnx
    device: CPU
    num_threads: 8

model:
  model_type: yolo-generic
  path: /config/helmet_head_person_s.onnx
  width: 320
  height: 320
  input_tensor: nchw
  input_pixel_format: rgb
  input_dtype: float
  labelmap_path: /config/labelmap.txt

objects:
  track:
    - person
  filters:
    person:
      min_area: 100       
      max_area: 999999
      threshold: 0.1     

# ✅ logger改成debug
logger:
  default: debug  # ⚠️ 改成debug！
  logs:
    frigate.object_detection: debug
    frigate.object_processing: debug
    detector.onnx_cpu: debug
    frigate.comms.mqtt: debug

He has been unable to recognize the person.
My labelmap.txt is:

person
helmet
head

My onnx model has tried code execution recognition:

What should I do?

[inkeep-ai-assistant[bot]]

Your custom ONNX model appears to work correctly outside Frigate (based on your test image), so the issue is likely in how Frigate is configured to use it. Here are the key areas to check:

1. Verify Model Configuration Matches Your ONNX Model

For YOLO models with the ONNX detector, your configuration should match what your model expects(1):

detectors:
  onnx_cpu:
    type: onnx

model:
  model_type: yolo-generic
  width: 320
  height: 320
  input_tensor: nchw
  input_dtype: float
  path: /config/helmet_head_person_s.onnx
  labelmap_path: /config/labelmap.txt

(1)

Important: Verify that your model actually expects input_pixel_format: rgb. Many YOLO models expect bgr format. Try changing:

model:
  input_pixel_format: bgr  # Try this instead of rgb

(1)

2. Check Your Labelmap Format

Your labelmap should have one label per line with the index matching your model's output classes(2). Since your model outputs person/helmet/head, your /config/labelmap.txt should be:

person
helmet
head

(2)

This looks correct, but ensure there are no extra spaces or blank lines.

3. Review Detection Thresholds

You've set extremely low thresholds for debugging, which is good:

objects:
  filters:
    person:
      threshold: 0.1
      min_area: 100

(3)

However, check the Frigate debug logs to see:

• If detections are happening but being filtered out
• What confidence scores your model is producing
• If there are any ONNX-specific errors

4. Common ONNX Model Issues

The ONNX detector supports YOLO models, but there are specific requirements(1):

• The model must output in a format Frigate can parse
• Input preprocessing (pixel format, normalization) must match what the model expects
• The model architecture must be compatible with Frigate's YOLO parser

Since your model works in your test code but not in Frigate, the most likely issues are:

1. Input pixel format mismatch (RGB vs BGR)
2. Input normalization - your input_dtype: float is correct, but verify the model expects values in [0, 1] range
3. Model output format - ensure your YOLO model outputs in a format Frigate expects

5. Enable Debug Logging

Your logger configuration is already set to debug, which is good(2). Check the Frigate logs specifically for:

frigate.object_detection
detector.onnx_cpu

These will show if the model is loading correctly and what detections (if any) are being produced.

Next Steps

1. Try changing input_pixel_format from rgb to bgr(1)
2. Check the Frigate debug logs for ONNX errors or detection output
3. Verify your ONNX model's expected input format matches your config
4. Temporarily remove all filters to see if raw detections are happening:

objects:
  track:
    - person
  # Remove filters temporarily for debugging

(3)

If you continue to have issues, share the relevant debug log output showing what happens when Frigate processes a frame with your model.

📚 Sources:

1. Supported Hardware
2. Advanced Options
3. Available Objects

---

Have another question? Just tag @inkeep.

[ZhaiSoul]

labelmap里label前面有数字id的，需要注意。参考该项目的labelmap：

Note that in the labelmap, labels with numeric IDs preceding them require special attention. Refer to the labelmap of this project for reference.

https://github.com/blakeblackshear/frigate/blob/dev/labelmap.txt

以及你应该想问的是person，但是你的翻译软件把他翻译为了people :(

同时顺带一提，detect里的长宽不需要和模型大小设置为一样，Frigate会裁剪画面发送给检测器。详细信息可以见中文文档：https://docs.frigate-cn.video/frigate/camera_setup#%E9%80%89%E6%8B%A9%E6%A3%80%E6%B5%8B%E5%88%86%E8%BE%A8%E7%8E%87

欢迎进入中文社群，我们可以提供中文支持

[13516513760]

我已经把他的labelmap.txt设置为

0  person
1  head
2  helmet

然后我的config.yml已经设置为

mqtt:
  enabled: true
  port: 1883
  topic_prefix: frigate
  client_id: emqx_MTc1Nj
  host: host.docker.internal
  user: admin
  password: znyj@123456

auth:
  enabled: true

cameras:
  name_of_your_camera:
    enabled: true
    ffmpeg:
      inputs:
        - path: rtsp://admin:[email protected]:554/h264/ch1/main/av_stream
          roles:
            - detect
    detect:
      enabled: true
      fps: 5
      # ✅ 移除width/height，使用全局的320x320
    motion: {}
    snapshots:
      enabled: true
      # ✅ 临时移除required_zones，让所有检测都能保存
      # required_zones:
      #   - person_test
      #   - user_test
    record:
      enabled: true
      retain:
        days: 7

    # ✅ 临时移除zones，稍后再添加
    # zones: {}
    review:
      detections:
        labels:
          - person

    mqtt:
      enabled: true
      timestamp: true
      bounding_box: true

    # ✅ 摄像头级别只追踪person
    objects:
      track:
        - person

    zones:
      person_area:
        coordinates: 
          0.315,0.466,0.791,0.398,0.985,0.539,0.983,0.85,0.079,0.948,0.072,0.548
        loitering_time: 0
        objects: person
version: 0.16-0

notifications:
  enabled: true
  email: [email protected]

# ✅ 全局detect配置
detect:
  fps: 1
  stationary:
    interval: 1
    threshold: 25
  enabled: true

snapshots:
  enabled: true
  bounding_box: true
  timestamp: true
  retain:
    default: 10

semantic_search:
  enabled: false
  reindex: false
  model_size: small

detectors:
  onnx_cpu:
    type: onnx
    device: CPU
    num_threads: 8

model:
  model_type: yolo-generic
  path: /config/helmet_head_person_s.onnx
  width: 320
  height: 320
  input_tensor: nchw
  input_pixel_format: bgr
  input_dtype: float
  labelmap_path: /config/labelmap.txt

# ✅ 全局objects配置
objects:
  track:
    - person

# ✅ logger改成debug
logger:
  default: debug  # ⚠️ 改成debug！
  logs:
    frigate.object_detection: debug
    frigate.object_processing: debug
    detector.onnx_cpu: debug
    frigate.comms.mqtt: debug

还是不行，请问我该怎么办？他依旧没有识别到person
还有，请问中文社群是什么？

[ZhaiSoul]

日志中是否有关于模型的信息？提供一下日志