Vehicle Detection and App Deployment with YOLOv8 (UA-DETRAC)

Overview

This project applies the YOLOv8 object detection framework to the UA-DETRAC dataset for vehicle detection, covering both model training and a lightweight desktop app deployment.

The complete pipeline includes:

• Data preprocessing
• Exploratory Data Analysis (EDA)
• Model training and optimization
• App deployment for quick video processing

---

Dataset

• UA-DETRAC: A large-scale real-world dataset for vehicle detection in traffic videos.
• Structure:
  • images/: Extracted video frames.
  • Test-Annotations/: Original XML labels.
  • labels/: YOLO-format labels generated in preprocessing.

---

Model Training Summary

• Framework: Ultralytics YOLOv8
• Steps:
  • XML annotations parsed and converted into YOLO [class, x_center, y_center, width, height] format.
  • Ignored regions are filtered.
  • Training/Validation/Test splits created.
• Evaluation Metrics:
  • mAP, Precision, Recall, F1-Score
• Optimization:
  • Hyperparameter tuning (batch size, learning rate, data augmentation like mosaic and HSV shift).
  • Post-processing predictions to remove overlaps and filter low-confidence detections.

---

Key Results

Version	Precision	Recall	F1-Score	F1 Variance
Before Optimization	0.6329	0.6390	0.6260	0.0213
After Filtering	0.7815	0.6172	0.6884	0.0166

---

Pipeline Details

1. Data Preprocessing

• Parse XML annotations (*.xml) and extract bounding boxes.
• Ignore regions marked in XML (<ignored_region> tag).
• Normalize bounding boxes into YOLO [class x_center y_center width height] format.
• Split dataset into train, val, and test sets.

2. Exploratory Data Analysis (EDA)

• Plot number of vehicles per frame.
• Visualize bounding box size and aspect ratio distribution.
• Heatmap of object locations to analyze camera angles and scene bias.
• Detect outliers in label data.

3. YOLO Model Training & Evaluation

• Model: YOLOv8 (Ultralytics).
• Trained using original images and generated YOLO labels.
• Evaluation Metrics:
  • mAP (mean Average Precision)
  • Precision, Recall
  • Loss curves (classification loss, localization loss, objectness loss).
• Save best checkpoint based on validation mAP.

4. YOLO Model Optimization

• Hyperparameter tuning:
  • Learning rate scheduling.
  • Anchor box optimization.
  • Data augmentation (mosaic, mixup, HSV shift).
• Retrain model with optimized settings.
• Re-evaluate mAP improvements.

5. Post Optimization

• Apply prediction filtering:
  • Remove predictions overlapping with ignored regions.
  • Filter small or low-confidence bounding boxes.
• Final evaluation on the cleaned test set.
• Save final prediction results.

App

The app supports:

• Selecting a video for frame-by-frame vehicle detection
• Exporting a new video with bounding boxes and labels
• Viewing processing progress with a progress bar
• Canceling detection if needed

---

Project Structure

Stage	File	Description
1	app.py	Main application source code
2	models/best.pt	Pretrained YOLOv8 model weights
3	requirements.txt	Python environment requirements

---

Environment Setup

1. Install dependencies

pip install -r requirements.txt

Required Python packages:

• ultralytics
• opencv-python
• PySide6
• tqdm
• torch (if not auto-installed)

---

How to Run

1. Launch the Application

After building the project with pyinstaller, go to the dist/app/ folder, where you will find app.exe. Make sure the models/ folder (containing best.pt) is placed inside dist/app/.

Double-click app.exe to launch the application.

2. User Interface

Button	Function
Select	Choose a video file for detection
Export	Generate processed video with bounding boxes
Cancel	Abort processing if needed

3. Processing Flow

• Select a video file (e.g., .mp4)
• The app extracts frames, runs YOLOv8 detection, and saves labels in temp/labels/
• After prediction, click Export to generate a new video with drawn bounding boxes and labels
• Output video is saved under runs/processed_videos/

4. Output Directory

Folder	Description
temp/frames/	Temporary extracted frames
temp/labels/	YOLO format detection results
runs/processed_videos/	Final annotated video output

---

Interface and Demos

Interface

Original Video Demo

Processed Video with Detections

---

Notes

• All temporary frames and labels will be automatically cleared after video export.
• If the video has different resolutions between frames, the app dynamically adapts frame size.
• Canceling during export stops the video generation safely.

---

Acknowledgements

• UA-DETRAC Dataset
• Ultralytics YOLOv8

---