# Forest Guard: Innovative AI-Powered App Monitoring Wood Transportation in Romania The Forest Guard project represents a pioneering approach to combatting illegal deforestation in Romania by harnessing the power of AI and computer vision technologies. It introduces a novel and innovative solution that empowers citizens and law enforcement agencies to monitor and track wood transportation activities effectively. The application employs advanced algorithms to detect license plate numbers using bounding box detection and OCR text recognition. It also extracts legal documents from the SUMAL 2.0 platform, ensuring compliance with forestry regulations and facilitating real-time monitoring of wood transports. Additionally, the app includes a state-of-the-art wood log detection feature, which utilizes a custom YOLOv8n model to detect and calculate wood volume accurately. By transforming pixels into centimeters and applying mathematical formulas, the app provides precise measurements, enabling users to verify the legality of wood transports. :::youtube[Title]{#ChgjLaMQUxc}

--DIVIDER--## 1. Background Deforestation in Romania is an alarming and rapidly escalating issue, posing a significant threat to the nation's environment, economy, and overall well-being. The statistics are striking: an estimated average loss of **3 hectares of forests every hour**, resulting in a substantial annual economic setback of approximately **$6 billion**, equivalent to a significant **2% of the country's GDP**. What's even more concerning is that **only 1% of illegal deforestation wood transport is currently caught**. Description

--DIVIDER--## 2. Objectives - Develop a mobile application that empowers citizens to report and verify timber transports, contributing to the reduction of illegal deforestation in Romania. - Provide law enforcement agencies with real-time, actionable data on potentially illegal timber transports, aiding in more effective monitoring and enforcement efforts. - Provide opportunities for team members to enhance their skills in AI, machine learning, project management, and collaboration through hands-on experience and peer learning in open source. - Create a wood volume measurement application for logging companies to accurately estimate wood volume and ensure sustainable harvesting practices. ![resized_objectives.png](resized_objectives.png)

--DIVIDER--## 3. Approach

A) License Plate Detection

Develop a bounding box detection model using YOLOv8 to identify license plates in images, incorporating a pixel-to-cm ratio to find a 52 cm reference length. Integrate an OCR API to recognize and extract text from the detected license plates, enabling the extraction of legal documents.

B) Extracting Documents

Extract from the national legal platform SUMAL to access real-time legal documents related to timber transports and forestry regulations.

C) Wood Log Detection

Implement a wood log detection model using YOLOv8 to identify and localize logs in images. Calculate wood log volume using mathematical techniques and compare it with real volume data to detect discrepancies and potential illegal logging activities.

--DIVIDER--## 4. Data

License Plate Dataset	Wood Log Dataset
Total Images: 1,457 Training: 1,075 Validation: 382 Sources: Roboflow Dataset (923 Images) Romanian License Plate Dataset (534 Images) Processing: Merged datasets Converted to YOLOv8 label format Image size: 640x640 No augmentations	Total Images: 2,585 Training: 1,846 Validation: 739 Sources: Roboflow Dataset (2,464 Images) HAWKWood Database (121 Images) Processing: Merged datasets Converted to YOLOv8 label format Image size: 640x640 No augmentations

License Plate Dataset

Wood Log Dataset

Total Images: 1,457

Training: 1,075

Validation: 382

Sources:

Processing:

Merged datasets
Converted to YOLOv8 label format
Image size: 640x640
No augmentations

Total Images: 2,585

Training: 1,846

Validation: 739

Sources:

Processing:

Merged datasets
Converted to YOLOv8 label format
Image size: 640x640
No augmentations

--DIVIDER--## 5. License Plate Recognition (Bbox + OCR)

A) Models

- YOLOv8 Nano (6.3 MB): Selected for its balance of small size and fast inference speed, making it suitable for mobile deployment.
- TrOCR: Transformer-based Optical Character Recognition with pre-trained models.

B) Experiments

-Several experiments were conducted with YOLOv8X and YOLO-NAS, but YOLOv8 Nano yielded the best results in terms of accuracy and efficiency.
-Created a 73-images evaluation dataset manually annotated to assess performance of different APIs and models. -Multiple experiments were conducted using various OCR techniques, including Pytesseract and TrOCR, with different sizes and fine-tuned on different datasets. TrOCR base-printed worked.

C) Hyperparameters

Parameter	Value
Model Architecture	YOLOv8N
Epochs	50
Learning Rate	0.002
Momentum	0.9
Loss	Sparse Cross Entropy
Optimizer	AdamW
Machine	Intel Core i7-1255U

D) Results and Evaluation

Metrics: Achieved mAP50 of 0.994 during testing.
Metrics

Examples: Model demonstrated robust performance under diverse lighting conditions.

--DIVIDER--## 6. SUMAL 2.0 Platform SUMAL 2.0 regulates and monitors the transportation of wood on public roads in Romania. It mandates a legal notice containing: - Legal volume. - License plate number. - Vehicle type. - Validity of the electronic notice. ![approachBd-removebg-preview.png](approachB.png)

--DIVIDER--## 7. Extracting Legal Documents ```python import requests from bs4 import BeautifulSoup license_plates = ['SB40DAP'] #-> License Plate Number base_url = 'https://inspectorulpadurii.ro/api/aviz' for plate in license_plates: response = requests.get(f'{base_url}/locations?nr={plate}').json() codes = response['codAviz'] if not codes: print("Legal Notice not found") for code in codes: resp_2 = requests.get(f'{base_url}/{code}').json() # Getting volume volume = resp_2['volum']['total'] # Getting valdity valid_from = get_date(resp_2['valabilitate']['emitere']) valid_to = get_date(resp_2['valabilitate']['finalizare']) print({'Code': code, 'Volume': volume, 'Validity': f'{valid_from} - {valid_to}'}) ```

--DIVIDER--## 8. Wood Log Detection (Multiple Bboxes)

A) Models

We opted for YOLOv8n due to its renowned performance in object detection tasks. Through training over 30 epochs, our model attained impressive accuracy, ensuring reliable detection of wood logs across various environmental conditions.

B) Experiments

Several experiments were conducted with YOLOv8S, DETR, and YOLO-NAS. YOLOv8 Nano yielded the best results in terms of accuracy and efficiency during testing and validation phases.

C) Hyperparameters

Parameter	Value
Model Architecture	YOLOv8N
Epochs	30
Learning Rate	0.01
Batch Size	8
Loss	VFL (CIOU+DFL)
Optimizer	AdamW
Machine	Intel Core i7-1255U

D) Results and Evaluation

Metrics: Achieved high accuracy in detecting wood logs under various conditions.
Metrics

Examples: Accurate bounding box predictions for logs with minimal false positives.

--DIVIDER--## 9. Pixel to Centimeter Ratio

Pixel to Centimeter Ratio	Diameter
Bounding Box Pixels : 52 (Standard EU License Plate Length)	The diameter is extracted by taking the length of the bounding box of each log and converting it into centimeters using the pixel-to-centimeter ratio.

--DIVIDER--## 11. Mobile App Development

A) Edge Detection for Real-Time Processing

Our application utilizes edge detection techniques to perform real-time processing directly on the user's device, without relying on cloud platforms or external APIs. This ensures smooth and fast performance even on mobile devices, providing users with a seamless experience.

B) Multiple Bounding Box Detection for License Plates

One of the key features of our app is its ability to detect multiple bounding boxes for license plates in real-time. This enables users to quickly identify and select the correct license plate number from the predictions. If the predicted number is incorrect, users can easily modify it manually by long-pressing on one of the options, ensuring accurate data capture.

C) Real-Time Wood Bounding Box Detection

In addition to license plate detection, our app also includes real-time detection of wood bounding boxes. Users can center the trucks within the frame, and the app will continuously detect and refine the bounding boxes until the user is satisfied with the results. This ensures accurate tracking and monitoring of wood transportation activities.

D) Manual Confirmation for Accuracy

To further enhance accuracy, our app waits for user confirmation before finalizing any predictions or calculations. This allows users to review and validate the detected information, ensuring that only accurate data is added to the verified trucks section. By empowering users to confirm the accuracy of the predictions, our app provides a reliable and trustworthy solution for wood transportation monitoring.

--DIVIDER--## 12. Demos + Try it Yourself

License Plate Recognition DEMO

:::youtube[Title]{#sdSSoybVFCA}

Wood Log Detection DEMO

:::youtube[Title]{#vqGdf4nYZzE}

Forest Guard DEMO

:::youtube[Title]{#imxzQmTLc1s}

Usage

To test the app, clone the repository and navigate to the `forest_guard` directory. Run the following command: ```python git clone https://github.com/patrick25076/forest_guard.git cd forest_guard flutter run ``` Ensure you are connected to an emulator or a real device. Contact me for the APK file if needed.

Contribution and Collaboration

I am open to collaboration and improvements. Feel free to reach out if you'd like to contribute or discuss potential enhancements. - Neicu Patrick - GitHub: [Profile](https://github.com/patrick25076) - Email: patrickneicu2006@gmail.com - LinkedIn: [Profile](https://www.linkedin.com/in/patrick-neicu-4bb567263/)