How to setup Smart NVR Pipeline to recognize automatic license plate ?

[topgun97]

I installed Intel Edge AI Box.I want to use it to recognize automatic license plate .My needs is：

1. Decodes 25 1080p H264 video channels from IPC
2. Stores video stream in my storage
3. OPen Model Zoo downloads license_plate_recognition_barrier_0007 .Runs it Inference on a subset of the video channels
4. Sends Inference metadata to my MQTT broker
5. Composes decoded video channels into multi-view 4K video wall stream
6. Encodes the multi-view video wall stream for remote viewing . Use VLC to Play it.
7. Display the multi-view video wall streams on attached monitor

How to do it?Thank for your help!

Regards,

Miao

[Peh_Intel]

Hi Miao,

Please refer to the Intel® Edge AI Box Article on How To Run Smart NVR Pipeline.

To use custom model (rather than default model) for running pipeline, please refer to the following steps. For more information, you can refer to custom_models/README.md.

1. Create a new directory under custom_models directory.

2. Put license_plate_recognition_barrier_0007 XML and BIN file in the new directory.

3. Create a model proc JSON file (Refer to license-plate-recognition-barrier-0007.json) in the new directory.

4. Update model XML file path and model proc JSON file path in config file (xx.cfg), you may only use the relative path under custom_models.

 For example, if you create a directory my_model for detection under custom_models directory, you can add the following path in smart_nvr.cfg file.

   ```

   DET_MODEL_NAME=my_model

   DET_MODEL_PATH=my_model/my_model.xml

   DET_MODEL_PROC_PATH=my_model/my_model_proc.json

Regards,

Peh

[topgun97]

Hi Peh，

            Thank you for your help, I refer to your steps, Omz_downloader and omz_converter models.Then create  model proc JSON file(license-plate-recognition-barrier-0007.json) under custom_models directory,  Smart NVR Pipeline works succeed, But the ratio of correct reads of license plate is too bad,  How can improve the accuracy?How to setup to  encode the multi-view video wall stream for remote PC viewing with VLC(url address). Thanks.

Regards,

Miao

[topgun97]

MQTTX recives messages:

{
"objects": [
{
"detection": {
"bounding_box": {
"x_max": 0.5104166666666666,
"x_min": 0.34375,
"y_max": 0.13920454545454544,
"y_min": 0.014204545454545454
},
"confidence": 1,
"label": "car",
"label_id": 2
},
"h": 44,
"id": 146,
"license_plate": {
"label": "<Guangdong>37",
"model": {
"name": "TensorFlow_Frontend_IR"
}
},
"region_id": 9555,
"roi_type": "car",
"w": 80,
"x": 165,
"y": 5
}
],
"resolution": {
"height": 352,
"width": 480
},
"source": "/home/aibox/sample_video/car_1080p20_180s_h264.mp4",
"timestamp": 73280000000
}

[Peh_Intel]

Hi Miao,

The accuracy is depends on what license plates the model has been trained on and many camera factors. For example, lighting, distance, moving objects and viewing angle. We recommend you to asses the environment and re-train model with own dataset.

Encode the multi-view video wall stream for remote PC viewing with VLC would be considered as a custom application. For any custom application, you can refer to the reference application of the documentation as the starting point. If there are any specific issues with the package we can assist.

Regards,

Peh

[topgun97]

Hi Peh,

   Thanks for the your reply .I will crefer to the reference application of the documentation .

Thanks & Regards,

Miao

[Peh_Intel]

Hi Miao,

Thank you for your question. This thread will no longer be monitored since we have provided answer and suggestion. If you need any additional information from Intel, please submit a new question. 

Regards,

Peh