Detection of stacked logs and Bulgarian license plate recognition

I created this small app to help process long video files in search of logging trucks. Ideally this would aid efforts to combat illegal logging. This repository includes two models which were fine-tuned by me with data collected from the public photo register of the South-Western State Enterprise. The fine-tuned models are:

• Object detection: YOLOv7 by Wong Kin Yiu.
• License plate detection IWPOD_NET by Claudio Jung.
• License plate OCR EasyOCR by JadedAI

Main features of the app include:

• Writing the output to a SQLite database
• Checking for new input from a GDrive folder
• The ability to organise detections in a group of frames
• Streamlit front-end for displaying the results

Installation

Install poetry:

pip install poetry

Clone the repository and navigate to the project directory:

poetry install

Configuration

Found in src/logtrucks/settings.yaml

Test configuration in tests/test_settings.yaml

# GDrive settings
enable_gdrive_source: False
gdrive_credentials: "credentials.json" # name of credential file stored in project root
gdrive_directory: "" # id of the shared GDrive directory
download_directory: "downloads" # name of directory for tmp downloads
# Detection logic
detection_max_frame_length: 50 # how many frames a single detection can span
detection_frame_buffer: 5 # number of frames until a detection with a new id is created
results_dir: "results"
write_to_db: True
rename_lp_image: True # add predicted license plate to name of file
# Models
yolo_weights: "v1.pt"
ocr_weights: "best_acc_300k_90percent"

Usage

Run cli examples:

poetry run logs -f tests/assets/127.jpeg

poetry run logs -f tests/assets/clip2.mp4

Results as seen in the Streamlit frontend:

Python:

from src.logtrucks.main import DetectLogic

dl = DetectLogic()
dl.start_detections("tests/assets/clip2.mp4")

Streamlit Frontend

A small streamlit app is included in the repo in order to visualise the contents of the SQLite database. The default db name is hardcoded in it. It will look for logtrucks.db in the project root and the default table name detections.

Run using:

poetry run streamlit run display_logsdb.py

Testing

Tests are run using pytest via poetry:

poetry run pytest

TODO

• Migrate IWPOD_NET from tensorflow to pytorch
• Refactor/clean up IWPOD_NET/OCR/YOLO code
• Investigate TensorRT for faster inference
• Complete readme
• Implement logs (as in monitoring)
• docstrings
• Improve error handling