/alpr-yolov8-python-ocr

Recognize license plates (and numbers) using fine-tuned #yolov8, #OCR (tesseract) and Hikvision camera

Primary LanguagePythonGNU Affero General Public License v3.0AGPL-3.0

Main Gate ALPR

Showcase

Showcase gif

(btw, the repository contains code and AI models for full reading without censoring. The video(s)/showcase(s)/image(s) are censored because I took data from production and don't want to dox people :D)

How does it work?

This is just a high-level explanation, if you want a more in-depth understanding, read the env config + source code :D

  1. When you start the web server, all env variables and ai models all loaded into memory.
  2. If you enabled result saving, the directory for results will get created
  3. Two threads get spin up,
    • First one reads frames from the IP cam / video and ensures to always be connected to the input source. If you're running in DEBUG, you'll see a window from the camera.
  4. The second one gets the latest frame and passes it to pure yolo.
  5. After analyzing, the program crops all cars into an array and checks whether you're not far enough
  6. After that, it passes the cropped image to fine-tuned yolo for license plates.
  7. The license plate gets cropped and pre-processed (more inside ./utils.py)
  8. Then the license plate gets separated into each character which gets passed to tesseract while using all possible threads
  9. License plate value gets finalized and validated (more on the validation)
  10. License plate and cropped car gets sent to all websocket-connected clients.
    • and optionally saved into DB/Results dir, based on your .env

How to setup

Prerequisites

  • python 3.11.2
  • docker (optional, in case you need db for local dev)
  • rust (optional, in case you want to try out the example provided client, however you can connect to the websocket server through any other app, like Postman)
  1. Go into ./ai/resources folder
  2. cat yolov8m_* > yolov8m.pt
cat yolov8l_* > yolov8l.pt
cat yolov8x_* > yolov8x.pt
cat andrewmvd_dataset_* > andrewmvd_dataset.zip
cat aslanahmedov_dataset_* > aslanahmedov_dataset.zip
cat tdiblik_lp_finetuned_yolov8m_* > tdiblik_lp_finetuned_yolov8m.pt
cat tdiblik_lp_finetuned_yolov8l_* > tdiblik_lp_finetuned_yolov8l.pt
cat tdiblik_lp_finetuned_yolov8x_* > tdiblik_lp_finetuned_yolov8x.pt
cp yolov8*.pt ..
  3. Go back into root folder
  4. pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
  5. Install tesseract
  6. pip install -r requirements.txt
  7. Make sure you're doing this with the same (unix) user the program is gonna be run with! In case you're doing this under another user, make sure to complete steps 4-6 with sudo -H -u <usr> pip3 install ... prefix.

Start the websocket and matching server

  1. Go into ./server
  2. Copy .env.development to .env
  3. Setup env
  4. Setup db server (optional, depends on your .env)
    1. Make sure you're following Microsoft's licensing requirements since the DB I chose is MSSQL. I did this because it's already integrated at my workplace. Personally, I use the docker only for development / testing.
    2. Go back into root folder.
    3. # (on Windows replace `pwd` with `pwd -W`)
docker run -e "ACCEPT_EULA=Y" -e "MSSQL_SA_PASSWORD=MyV€ryStr0ngP4ssW0rĐ" \
 -p 1433:1433 --name main_gate_aplr_db --hostname main_gate_aplr_db \
 -v $(pwd)/db/data:/var/opt/mssql/data \
 -v $(pwd)/db/log:/var/opt/mssql/log \
 -v $(pwd)/db/secrets:/var/opt/mssql/secrets \
 -d mcr.microsoft.com/mssql/server:2022-latest
    4. docker exec -it main_gate_aplr_db "bash"
    5. export QUERY_TO_EXECUTE="{CONTENTS_OF_./db/init.sql}"
    6. /opt/mssql-tools/bin/sqlcmd -S localhost -U SA -P "MyV€ryStr0ngP4ssW0rĐ" -Q "$QUERY_TO_EXECUTE"
  5. Automate the startup process (optional, recommended)
    • Create a file named main-gate-alpr-server.service inside /etc/systemd/system
    • Copy contents of the ./server/main-gate-alpr-server.example.service into main-gate-alpr-server.service and make sure to fill in the <<usr>> in the process.
    • sudo systemctl daemon-reload
    • sudo systemctl enable main-gate-alpr-server.service
    • sudo systemctl restart main-gate-alpr-server.service
    • sudo systemctl status main-gate-alpr-server.service (optional, make sure the process is running)
    • sudo journalctl -u main-gate-alpr-server.service (optional, see logs)
  6. python server.py

Start the example client (optional)

  1. Go into ./client
  2. Make sure variable WEBSOCKET_URL is set up correctly inside ./client/src/main.rs based on your server's .env (sorry, I was too lazy to implement env config only because of this one variable)
  3. Run cargo r
  4. You should see one of the following states
    • Client can be trying to connect
      • reconnecting picture
    • Client can be disconnected (hover over for more info)
      • connection closed picture
    • Or working as intended! (if you setup everything correctlly)
      • connected and working

Train your own model (optional)

  1. Go into ./ai folder
  2. export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512 (on Linux) or set PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512 (on Windows)
  3. python prepare.py
  4. Go into train.py and configure which pre-trained model you want to use.
  5. python train.py
    • If you encounter an error with the path, running the program again will probably solve it. For some reason (ultralytics related ¯_(ツ)_/¯) it sometimes fails on the first launch in a new directory. If your error persists, open a new issue.
    • If you run out of GPU memory, or encounter any error while training, play around with the batch variable. On my RTX 3070 Ti, 16 batches for smaller models (n/s/m) is enough, however for larger models (l/x), I had put the batches down to 8

Test your/provided models visually (optional)

  1. Go into ./ai folder
  2. Run the following command: python test.py {path_to_your_or_provided_model} {path_to_image_to_test_on}
  3. You can now visually check your AI model's output (you should probably use your own image for testing) example model test output

How to configure env

You can find an example config at ./server/.env.development

Base (change behavior in a noticable way)

  • DEBUG
    • If you want to see debug information, set the value to True.
    • This will not only show logs, open a window where you can see the video feed, but also save intermediate files while matching, so you can inspect them, into ./server/intermediate_detection_files
    • If you want to disable this, just remove the option or set it to any other value than True
  • WS_PORT
    • required
    • Determines the port of the web socket server.
  • RTSP_CAPTURE_CONFIG
    • required
    • Video input for matching.
    • This can be either a custom video, in that case, specify the file path (eg. "./test.mp4"), or RTSP path (eq. "rtsp://{username}:{password}@{ip}:{port_probably_554}/Streaming/channels/1/"). You can probably find RTSP config info at the following path of your IP cam (http://{ip}:{port_probably_80}/Streaming/channels/1/). If you're unable to find the config, try changing around the channels part of the url (eq. 1/2/3,...).
    • If you're not sure whether you've configured this variable correctly, go into ./server and run python test_rtsp.py. It will use your .env, just as the web server would, and provide you with visual feedback.
  • PURE_YOLO_MODEL_PATH
    • required
    • Pure unedited yolo(v8) model is used for car recognition.
    • As long as it's yolov8, it does not matter, which type (n/s/m/l/x) you choose. To save you some time, you can find the models at ./ai/resources/yolov8*.pt.
    • Example value could like "../ai/resources/yolov8n.pt"
  • LICENSE_PLATE_YOLO_MODEL_PATH
    • required
    • Fine-tuned model path used for matching license plates.
    • I have fine-tuned every single version of Yolo already, so you don't have to. You can find it under tdiblik_lp_finetuned_yolov8*.pt (where * represents the model type)
  • DB_ENABLED
    • If you want to insert results into the database, set this value to True.
    • If you set this value to True, make sure to follow the db-setup guide.
    • You can find the db schema at ./db/init.sql. Only id, license_plate, and captured_at fields are required for the web socket server to work, however since this server is used as a part of a proprietary solution, I just included the whole schema as is in the proprietary solution :D. Feel free to edit it to fit your use case.
    • if you want to disable this, just remove the option or set it to any other value than True
    • If you enable DB, all of the following env options are considered as required
    • DB_SERVER
      • Sql server IP / DNS alias (eg. localhost)
    • DB_PORT
      • You probably want to set the value to 1433 since that is the default MSSQL port.
    • DB_NAME
      • Name of the database to use inside your database server
      • If you used the ./db/init.sql, you want to set this value to lpdb
    • DB_USER
      • Database user to use
      • If you used the docker setup command, or didn't change defaults (not recommended), you want to set this value to SA
    • DB_PASSWORD
      • Password of the database user
      • If you used the docker setup command, you want to set this value to MyV€ryStr0ngP4ssW0rĐ
  • SAVE_RESULTS_ENABLED
    • If you want to save car and license plate images matched, set this value to True.
    • If you want to disable this, just remove the option or set it to any other value than True
    • If you enable saving results, all of the following env options are considered as required
    • RESULTS_PATH
      • Path where to save results (will be created if not exists)
      • Eg. "./results"
  • SHOULD_SEND_SAME_RESULTS
    • Once a license plate is matched, there's a check on whether to send the result. Basically, if the program already sent the license plate value, same as the one currently matched, in the last 5 minutes, the license plate matched will get ignored.
    • If you want to enable this behavior (default), just remove the option or set it to any other value than True
    • If you want to disable this behavior and process all matches, set this value to True.

Custom tweaks (tinkering with these can become a silent problem if you don't know what you're doing)

  • SHOULD_TRY_LP_CROP
    • Some models may have problems where they recognize the part next to the license plate as a part of the license plate (see picture below)
    • If you encounter this behavior from your models, set this value to True and this will try to programatically cut that part out.
    • Otherwise, just remove the option or set it to any other value than False
    • example of that annoying part getting recognized
  • MINIMUM_NUMBER_OF_CHARS_FOR_MATCH
    • You probably want to set this value to default 4 (shortest license plate in Europe)
    • Basically, if you match a license plate, however, you're unable to find more than MINIMUM_NUMBER_OF_CHARS_FOR_MATCH of contours (pre-precessed, cropped, single characters of license plate), there is no reason to try reading it (and waste CPU time/cycles) since the license plates won't be valid anyway. So, the smart decision is just to skip it.
    • If you didn't understand the previous sentence, but want to tinker with it anyway, just set it to the minimum number of characters a license plate has to contain for it to be a match.
  • NUMBER_OF_VALIDATION_ROUNDS
    • You probably want to set this value to default 3
    • So, the quality of the camera can be bad and the AI's don't work correctly 100% of the time. That's why there's the NUMBER_OF_VALIDATION_ROUNDS variable. Basically, the entire matching process output is saved to an intermediate array that gets filtered (filtering affected by NUMBER_OF_OCCURRENCES_TO_BE_VALID) every Nth run, and after that, the results are sent/saved/etc. If you want to see how the filtering works, check out ./server/server.py, function validate_results_between_rounds
    • Also, if you set NUMBER_OF_OCCURRENCES_TO_BE_VALID > NUMBER_OF_VALIDATION_ROUNDS, then you will never match any license plates :D
  • NUMBER_OF_OCCURRENCES_TO_BE_VALID
    • You probably want to set this value to default 2, or, if you tinkered with NUMBER_OF_VALIDATION_ROUNDS, you probably want it to be NUMBER_OF_VALIDATION_ROUNDS - 1
    • Inside the intermediate array, how many times the license plate number has to occur for it to be considered valid and pass validation. If you want to see how the filtering works, check out ./server/server.py, function validate_results_between_rounds
  • SKIP_BEFORE_Y_MAX
    • This is a tricky one, however, probably the most important one, while tinkering. Ok, when we find a car in a picture, we can assume that if it's too far from the top of the picture, the license plate is not readable, hence we can ignore that car and not waste CPU time/cycles.
    • ymax is the bottom line of the matched car.
    • Let's see this in practice, in the following picture, none of the cars will get matches, because we can assume that the car is far enough, for the license plate to be messy and not readable. Picture where both cars are far in this picture, both cars are far and we can assume that reading the license plate will be a waste of time.
    • However, in the following picture, one of the cars is close enough for the license plate to be readable, and therefore it's not ignored. Picture where one car's license plate is readable
    • This optimization enables us to only run car detection and optionally skip license plate detection and OCR phase. This speeds up the program A LOT and even provides better results because we ignore "trash" data.

Development Notes

  • When adding stuff into resources, for anything over 25MB, use the following command: split -b 25M --numeric-suffixes <name> <name>_ and add proper documentation on how to build it back together after clone.
  • If you see lp alias inside the code, it's probably a shorthand form of license plate
  • At the moment, the biggest problem is the OCR engine itself, because sometimes, it does not behave correctly. I have considered using PaddleOCR, however, I was unable to get it working :(. Nevertheless, it works well enough in production (missing only like 1 in 20 cars) for me to not care xdd. If somebody decides to implement other OCR (or train a custom one 👀), I'll be more than happy to merge that PR.

Acknowledgements

./ai/resources/yolov8*

Yolo models were downloaded from the ultralytics repository. I was unable to find any documentation on how to credit them, please, if you do, please send a pull request.

./ai/resources/andrewmvd_dataset.zip

BibTeX
@misc{make ml},
title={Car License Plates Dataset},
url={https://makeml.app/datasets/cars-license-plates},
journal={Make ML}

I downloaded the following kaggle dataset.

./ai/resources/aslanahmedov_dataset.zip

I downloaded the following kaggle dataset. The original author is ASLAN AHMEDOV, however the sources cited are "web scraping"

Other