Real-Time Vehicle Tracking & Counting System

A comprehensive real-time vehicle tracking and counting system that combines optical flow and YOLOv8 deep learning detection for accurate vehicle tracking and counting. Features a modern web-based interface built with Streamlit for easy video processing and analysis.

Developed as part of CPS843 - Introduction to Computer Vision at Toronto Metropolitan University. Formerly known as RyersonU.

---

🎯 What It Does

This application tracks and counts vehicles in video footage using advanced computer vision techniques. It can:

• Detect vehicles using either optical flow (Lucas-Kanade) or YOLOv8 deep learning
• Track vehicles across frames with persistent track IDs
• Count vehicles as they cross user-defined regions of interest (ROI)
• Visualize results with annotated videos showing track IDs and trajectories
• Provide statistics including total counts, directional counts, and real-time analytics

---

📸 Application Interface

Main Dashboard

The web interface provides an intuitive dashboard with:

• Video Input Selection: Choose from demo videos or upload your own MP4/MOV files
• Processing Parameters: Adjust YOLO confidence threshold, minimum box size, and detection methods
• ROI Configuration: Set up counting lines or polygons with auto-detection support
• Real-Time Statistics: View vehicle counts, directional statistics, and processing progress

Processing Interface

The system processes videos in chunks, providing real-time feedback and progress updates. You can adjust confidence thresholds to balance between detection accuracy and false positive rates.

Results Visualization

Processed videos show:

• Track IDs: Each vehicle is assigned a unique ID (e.g., ID0, ID1, ID2)
• Trajectory Paths: Visual trails showing vehicle movement
• Bounding Boxes: Detection boxes around each vehicle
• Counting Lines: Visual representation of ROI boundaries

Multiple Line Configuration

Advanced users can configure multiple counting lines for enhanced accuracy and validation. Each vehicle is counted only once, even when crossing multiple lines.

---

✨ Key Features

• Dual Detection Methods:

  • Optical Flow (Lucas-Kanade) for fast, efficient tracking
  • YOLOv8 deep learning for high-accuracy detection

• Automatic ROI Detection: Uses YOLOv8 to analyze vehicle movement and suggest optimal counting lines

• Persistent Track IDs: Each vehicle maintains a unique ID throughout the video, preventing double-counting

• Chunk-Based Processing: Processes videos in frames for faster analysis and real-time feedback

• Flexible ROI Types:

  • Line ROI: Count vehicles crossing a line
  • Polygon ROI: Count vehicles entering/exiting a region

• Real-Time Analytics: Live statistics showing total counts, directional counts, and processing progress

• Web-Based Interface: Easy-to-use Streamlit dashboard with drag-and-drop video upload

---

🚀 Quick Start

1. Installation

cd Project
pip install -r requirements.txt

2. Run the Web Interface

streamlit run app.py

The application will open in your browser at http://localhost:8501

3. Process a Video

1. Select a demo video or upload your own MP4/MOV file
2. Choose detection method (YOLOv8 recommended for accuracy)
3. Adjust confidence threshold (0.4 recommended for best balance)
4. Set up ROI (use auto-detect or draw manually)
5. Click "PROCESS VIDEO" and wait for results

---

🔧 How It Works

Optical Flow Tracking

Uses Lucas-Kanade sparse optical flow to track feature points across frames. Detects "good features to track" using Shi-Tomasi corner detection and tracks them frame-to-frame. Fast and efficient, achieving 60+ FPS on CPU.

YOLOv8 Detection

Deep learning-based object detection using YOLOv8 model. Provides superior accuracy (90-95% at confidence threshold 0.4) by detecting vehicles as complete objects rather than tracking features. Operates at 10-15 FPS on CPU.

Kalman Filtering

Smooths trajectories and predicts positions during brief occlusions using a constant velocity model. Helps maintain track continuity even when vehicles are temporarily hidden.

ROI-Based Counting

Counts vehicles when their track IDs cross defined boundaries (lines or polygons). Each track ID can only trigger one count, preventing double-counting.

---

📊 Performance

• Accuracy: 90-95% vehicle counting accuracy at confidence threshold 0.4
• Speed:
  • YOLOv8: 10-15 FPS on CPU
  • Optical Flow: 60+ FPS on CPU
• Detection Classes: Car, Motorcycle, Bus, Truck (from COCO dataset)

---

🛠️ Tech Stack

Technology	Purpose
OpenCV	Video processing, optical flow, visualization
YOLOv8 (Ultralytics)	Deep learning-based vehicle detection
Streamlit	Web-based user interface
FilterPy	Kalman filter implementation
NumPy/SciPy	Numerical computations
Plotly	Interactive charts and statistics

---

📁 Project Structure

real-time-vehicle-tracking-cv/
├── Project/                    # Main application
│   ├── app.py                  # Streamlit web interface
│   ├── main.py                 # Command-line interface
│   ├── src/                    # Core modules
│   │   ├── video_processor.py
│   │   ├── optical_flow_tracker.py
│   │   ├── yolo_detector.py
│   │   ├── vehicle_counter.py
│   │   └── utils.py
│   ├── data/                   # Sample videos
│   └── output/                 # Processed videos
└── README.md                   # This file

---

👥 Author

Arshia Rahim
Computer Engineering (Software) @ Toronto Metropolitan University
GitHub: @ArshiaRx

Collaborators: Ansugan Subramaniam, Wajeehul Hassan

---

📚 Course Information

CPS843 - Introduction to Computer Vision
Fall 2025 • Toronto Metropolitan University

---

🔗 Repository

GitHub: https://github.com/ArshiaRx/real-time-vehicle-tracking-cv.git

---

📖 Documentation

For detailed technical documentation, usage examples, and configuration options, see Project/README.md.

---

📄 License

Educational project. Feel free to reference with attribution.