Description of the FALPRS project

This project is a replacement for the old one. Main differences:

• PostgreSQL is used as a DBMS.
• The project uses userver, an open-source asynchronous framework.
• Strict adherence to data types in API requests. For example, if a numeric field is expected, it cannot be enclosed in quotation marks.
• Added a license plate recognition system (LPRS).

Content

• LPRS
  • Neural network models used
  • General scheme of interaction with LPRS
  • Automatic number ban
• FRS
  • Neural network models used
  • General scheme of interaction with FRS
• Building and setting up the project
  • System requirements
  • Installing NVIDIA drivers
  • Installing Docker Engine
  • Installing NVIDIA Container Toolkit
  • Installing PostgreSQL
  • Building the project
  • Creating TensorRT neural network model plans
  • Project configuration
• Examples
  • LPRS
  • FRS
• Tests
  • LPRS
  • FRS
• Importing data from an old FRS project

LPRS

The system is designed to recognize license plates and detect special vehicles with flashing lights: ambulance, emergency, police, etc. Currently, Russian registration plates of type 1 (ГОСТ Р 50577-93) and 1А (ГОСТ Р 50577-2018) are supported:

Neural network models used

The service works with four neural networks: VDNet, VCNet, LPDNet and LPRNet. VDNet is designed to search for vehicles using a received image from a video camera. VCNet determines whether each vehicle found is a special one. LPDNet is designed to search for license plates. LPRNet is designed to recognize license plates from data received by LPDNet. VDNet, LPDNet and LPRNet models are trained using Ultralytics. The VCNet model was obtained by transfer learning with fine-tuning. Based on this model. NVIDIA Triton Inference Server is used for inference of neural networks.

General scheme of interaction with LPRS

To interact with the service, an API is used, a detailed description of which is in the repository in the file docs/openapi.yaml. First you need to register the video streams that the system will work with. The API method addStream is used for this. Main parameters:

• streamId - internal for your backend (external for LPRS) video stream identifier;
• config - configuration parameters. Example request body:

{
  "streamId": "1234",
  "config": {
    "screenshot-url": "https://my.video.stream/capture",
    "callback-url": "https://my.host/callback?streamId=1"
  }
}

Now LPRS knows that frames should be captured using the specified screenshot-url parameter, and when a special vehicle is detected or license plate recognition is detected, a short HTTP information should be sent using the POST method to callback-url. In order for LPRS to process frames, you need to call the startWorkflow API method. Example request body:

{
  "streamId": "1234"
}

The system will begin a cyclical process: receiving a frame, processing it using neural networks, sending, if required, information to callback-url. Pause for a while and again receive the frame, process it, etc. To stop the process, you need to call the stopWorkflow API method. Example request body:

{
  "streamId": "1234"
}

To reduce the load on the system, we recommend calling startWorkflow and stopWorkflow, for example, according to camera video analytics: motion detection, line crossing, intrusion detection, etc. When receiving brief information about recognized license plates, your backend must process it. For example, check the received numbers with the list of allowed ones and open the gate or simply ignore them. License plate recognition events are stored in LPRS for some time. If complete information about an event is required, then you need to use the getEventData API method. To remove a video stream, use the removeStream API method. Using listStreams you can get information about video streams registered in LPRS.

Automatic number ban

To prevent spam requests to callback-url, a two-stage number ban is used. If the system sees the number for the first time, then after processing it will be at the first stage of the ban: for some time (configuration parameter ban-duration), the number is ignored, regardless of its location in the frame. If the system sees this number again at the first stage, the ban time will be extended. At the end of the first stage, the number will be included in the next one. At the second stage of the ban (configuration parameter ban-duration-area), the number is ignored until it changes its location in the frame. When changing the location (configuration parameter ban-iou-threshold), the number will be processed and will again go to the first stage. At the end of the second stage, the number is removed from the ban. If you do not want automatic banning of numbers to be applied, then you need to set the value of the ban-duration or ban-duration-area configuration parameter to a zero value, for example 0s (0 seconds).

FRS

The system is designed for facial recognition.

Neural network models used

At the moment, FRS uses three models in its work:

• scrfd - designed to search for faces in an image. Link to the project.
• genet - designed to determine whether a mask or dark glasses is on the face. Based on this project. The model was obtained by transfer learning with fine-tuning on three classes: an open face, a face in a mask, a face in dark glasses.
• arcface - designed to calculate a biometric face template. Link to the project.

General scheme of interaction with FRS

To interact with the service, an API is used, a detailed description of which is in the repository in the file docs/openapi.yaml. First of all, your backend registers video streams using the addStream API call. The main parameters of the method are:

• streamId - internal for the backend (external for FRS) identifier of the video stream;
• url is the URL for capturing a frame from the video stream. FRS does not decode video, but works with individual frames (screenshots). For example, the URL might look like http://hostname/cgi-bin/images_cgi?channel=0&user=admin&pwd=password
• callback is the URL that FRS will use when recognizing a registered person. For example, http://backend-address/face_recognized?stream_id=1

To register faces, the registerFace method is used. Main parameters of the method:

• streamId - video stream identifier;
• url is the URL of the face you want to register. For example, http://backend-address/image_to_register.jpg In case of successful registration, the internal for FRS (external for the backend) unique identifier of the registered person - faceId - is returned.

To start and end frame processing, the motionDetection method is used. The main idea is for FRS to process frames only when motion is detected in front of the video camera. Method parameters:

• streamId - video stream identifier;
• start - sign of the beginning or end of movement. If start=true, then FRS starts processing a frame from the video stream every second (set by the delay-between-frames parameter). If start=false, then FRS stops processing.

Frame processing means the following chain of actions:

1. Finding faces using the scrfd neural network.
2. If faces are found, then each one is checked for “blur” and “frontality”.
3. If the face is not blurred (clear image) and is frontal, then the presence of a mask and dark glasses is determined using the genet neural network.
4. For each face without a mask and without dark glasses, using the arcface neural network, a biometric face template, also known as a descriptor, is calculated. Mathematically, a descriptor is a 512-dimensional vector.
5. Next, each such descriptor is compared in pairs with those registered in the system. The comparison is made by calculating the cosine of the angle between the descriptor vectors: the closer the value is to one, the more similar the face is to the registered one. If the maximum cosine value is greater than the threshold value (tolerance parameter), then the face is considered recognized and FRS calls a callback (face recognition event) and specifies the descriptor identifier (faceId) and the internal identifier for FRS (external for the backend) as parameters events (eventId). If there are several recognized faces in one frame, then the callback will be called for the highest quality ("best") face (blur parameter).
6. Every unblurred, frontal, maskless, sunglasses-free, best face found is temporarily stored in the FRS log.

Using the bestQuality method, you can request the “best” frame from the log from FRS. For example, a person unfamiliar to the system approached the intercom and opened the door with a key. The backend knows the opening time with the key (date) and requests the best frame from FRS. FRS looks in its log for a frame with the maximum blur from the time range [date - best-quality-interval-before; date + best-quality-interval-after] and produces it as a result. Such a frame is a good candidate for registering a face using the registerFace method. As a rule, for good recognition it is necessary to register several faces of one person, including from frames taken in the dark, when the camera switches to infrared operating mode.

Building and setting up the project

System requirements

• NVIDIA GPU with Compute Capability greater than or equal to 6.0 and 4 GB memory or more. Details can be found, for example, here.

• DBMS PostgreSQL 14 or higher.

To get the source code you need git. If not installed, then run the command:

sudo apt-get install -y git

Getting the project source code:

cd ~
git clone --recurse-submodules https://github.com/rosteleset/falprs.git

Installing NVIDIA drivers

If the system already has the latest NVIDIA drivers installed, skip this step. To install, you can use the script scripts/setup_nvidia_drivers.sh. Basic commands taken from here.

sudo ~/falprs/scripts/setup_nvidia_drivers.sh

After installation, you need to restart the operating system:

sudo reboot

Installing Docker Engine

If Docker Engine is already installed on the system, then skip this step. To install, you can use the script scripts/setup_docker.sh. Basic commands taken from here.

sudo ~/falprs/scripts/setup_docker.sh

Installing NVIDIA Container Toolkit

If NVIDIA Container Toolkit is already installed on the system, then skip this step. To install, you can use the script scripts/setup_nvidia_container_toolkit.sh. Basic commands taken from here.

sudo ~/falprs/scripts/setup_nvidia_container_toolkit.sh

After installation, you are prompted to restart docker:

sudo systemctl restart docker

Установка PostgreSQL

If PostgreSQL is not installed, then run the command:

sudo apt-get install -y postgresql

Run psql:

sudo -u postgres psql

Run the SQL commands, specifying your password instead of "123":

drop user if exists falprs;
create user falprs with encrypted password '123';
create database frs;
grant all on database frs to falprs;
alter database frs owner to falprs;
create database lprs;
grant all on database lprs to falprs;
alter database lprs owner to falprs;
\q

Building the project

To build the project, you can use the script scripts/build_falprs.sh. The project is built using LLVM. The version is set by the LLVM_VERSION variable. The major PostgreSQL version is set by the PG_VERSION variable. You can find out the installed version of PostgreSQL with the command:

psql --version

The project's working directory is specified by the FALPRS_WORKDIR variable (default value /opt/falprs), the version of the container with Triton Inference Server is specified by the TRITON_VERSION variable.

Compute Capability and Latest Container Version Support Matrix

Compute Capability	GPU Architecture	Container Version	TensorRT
6.x	Pascal	24.04	8.6.3
7.0	Volta	24.09	10.4.0.26

If you have a GPU with Compute Capability 7.5 or higher, you can use the latest version of the container. For example, build for Ubuntu 24.04:

sudo LLVM_VERSION=18 PG_VERSION=16 TRITON_VERSION=24.09 ~/falprs/scripts/build_falprs.sh

For Ubuntu 22.04:

sudo LLVM_VERSION=15 PG_VERSION=14 TRITON_VERSION=24.09 ~/falprs/scripts/build_falprs.sh

Creating TensorRT neural network model plans

For inference, TensorRT plans are used, which can be obtained from neural network models in the ONNX (Open Neural Network Exchange) format. To create plans in the working directory, you can use the script scripts/tensorrt_plans.sh. The project's working directory is specified by the FALPRS_WORKDIR variable (default value /opt/falprs), the Triton Inference Server version is specified by the TRITON_VERSION variable:

sudo TRITON_VERSION=24.09 ~/falprs/scripts/tensorrt_plans.sh

Project configuration

To initially populate the databases, run the commands, specifying the values ​​of the variables with the prefix pg_ (replace the password “123” with the one you specified when creating the PostgreSQL user):

pg_user=falprs pg_passwd=123 pg_host=localhost pg_port=5432 pg_db=frs ~/falprs/scripts/sql_frs.sh
pg_user=falprs pg_passwd=123 pg_host=localhost pg_port=5432 pg_db=lprs ~/falprs/scripts/sql_lprs.sh

The project configuration is in the file /opt/falprs/config.yaml The main parameters have a description in the comments. Some values ​​need to be replaced.

• In the sections components_manager -> components -> lprs-postgresql-database and components_manager -> components -> frs-postgresql-database in the values ​​of the dbconnection attributes, replace the password "123" with the one you specified when creating the PostgreSQL user, as well as other access details, if they differ.
• In the components_manager -> task_processors -> main-task-processor section, replace the worker_threads value with the number of CPU cores of your server. The quantity can be viewed, for example, using the command:

cat /proc/cpuinfo | grep processor | wc -l

• In the components_manager -> task_processors -> fs-task-processor section, replace the worker_threads value with the number of CPUs on your server multiplied by 4.

To start a container with Triton Inference Server run the command:

sudo TRITON_VERSION=24.09 ~/falprs/scripts/triton_service.sh

To create the falprs service and rotate logs, run the command:

sudo ~/falprs/scripts/falprs_service.sh

Examples

Installing dependencies:

sudo apt-get install -y nodejs npm libcairo2-dev libpango1.0-dev libjpeg-dev libgif-dev librsvg2-dev

LPRS

Go to the directory with examples:

cd ~/falprs/examples/lprs

Installing dependencies:

npm i

Launch a test service to process callbacks:

node lprs_backend.js

Add a “test” video stream. In the new console:

cd ~/falprs/examples/lprs
./test_add_stream.sh

Check that the stream is in the database:

./test_list_streams.sh | jq

The response must contain a json file. Wait a minute for new data to appear in the FALPRS cache, execute the command:

./test_workflow.sh

As a result of executing this command in the console with the test service lprs_backend.js, we should see lines like:

[2024-10-16 12:37:16.108] Callback: {"streamId":"test001","date":"2024-10-16T09:37:16.103891056+00:00","eventId":1,"plates":["O588OA68"],"hasSpecial":false}
[2024-10-16 12:37:16.108] Matched numbers: O588OA68
[2024-10-16 12:37:16.162] Save image to file: /tmp/lprs_backend/screenshots/49b507ac-9b7d-43a6-94f0-18f647fc1f5c.jpg

And in the directory /tmp/lprs_backend/screenshots a file with an image should appear:

Deleting the “test” video stream:

./test_remove_stream.sh

Go to the console from which we launched the test service to process callbacks and stop it.

FRS

Go to the directory with examples:

cd ~/falprs/examples/frs

Installing dependencies:

npm i

Launch a test service to process callbacks:

node frs_backend.js

Add a “test” video stream. In the new console:

cd ~/falprs/examples/frs
./test_add_stream.sh

Check that the stream is in the database:

./test_list_streams.sh | jq

The response must contain a json file. Wait 10 seconds for the new data to enter the FALPRS cache, and execute the face registration command:

./test_register_face.sh

Check that the person is in the database:

./test_list_all_faces.sh | jq

The response must contain a json file. Wait 10 seconds for the new data to enter the FALPRS cache, execute the command:

./test_workflow.sh

As a result of executing this command in the console with the test service frs_backend.js, we should see lines like:

[2024-10-21 15:34:29.180] Callback: {"faceId":1,"eventId":1}
[2024-10-21 15:34:29.180] This is Albert Einstein
[2024-10-21 15:34:29.220] Save image to file: /tmp/frs_backend/screenshots/492d7722d038434d9e2676648991e65e.jpg

And in the directory /tmp/frs_backend/screenshots a file with an image should appear:

Tests

Installing dependencies:

sudo apt-get install -y python3-requests python3-pytest python3-pytest-order

Creating directories and test databases:

cd ~/falprs/tests
./test_prepare.sh

LPRS

pytest -v -s test_api_lprs.py

FRS

pytest -v -s test_api_frs.py

After finishing the tests, delete the directories and test databases:

./test_clean.sh

Importing data from an old FRS project

To copy data from an old project to a new one, you can use the import_data.py script from the utils directory of this repository. Before running the script, make sure that your old service is stopped, and that the new one is correctly configured and stopped. Installing dependencies:

sudo apt-get install -y pip python3-virtualenv

Go to the directory with the script:

cd ~/falprs/utils

Copying the file:

cp config.sample.py config.py

In the config.py file, replace the variable values ​​according to your configuration of the old and new service. Variables of the form mysql_* and *_old refer to the old project, pg_* and *_new - to the new one. Run the commands:

virtualenv venv
source venv/bin/activate
pip install -r requirements.txt
python import_data.py

After the script runs, a success message should appear. Delete the virtual environment:

deactivate
rm -rf ./__pycache__/
rm -rf venv