🎥 Video Inventory Detection System

An optimized object detection and tracking system for counting inventory items in videos using YOLO models. Features advanced optimizations for 29x faster processing on Apple Silicon.

🚀 Features

• Real-time Object Detection - Detects and counts objects in video frames
• Object Tracking - Tracks unique objects to avoid duplicate counts
• Multi-Model Support - Supports grocery_yolov5s.pt, YOLOv8n, ONNX, and OpenVINO formats
• High Performance - 29x faster than baseline with optimizations
• Automatic Fallback - Intelligently falls back to available models

📊 Performance

Configuration	Speed per Frame	Speedup
Original (YOLOv8x @ 640px)	~700ms	1x
Optimized (YOLOv8n @ 320px)	~24ms	29x faster ⚡

Optimizations Applied:

• ✅ Reduced image size from 640px to 320px (3x speedup)
• ✅ Using YOLOv8n instead of YOLOv8x (smaller, faster model)
• ✅ ONNX Runtime support (3x-6x faster on some systems)
• ✅ OpenVINO support (5x-10x faster on Intel CPUs)
• ✅ Optimized for Apple Silicon (M1/M2/M3)

📁 Project Structure

Video_model/
├── src/
│   ├── detect_inventory.py    # Frame-based object detection
│   ├── track_inventory.py     # Video-based object tracking
│   ├── extract_frames.py      # Video frame extraction
│   ├── run.py                 # Main execution script
│   └── utils.py               # Utility functions
├── videos/
│   └── sample.mp4             # Input video files
├── output/
│   └── inventory.json         # Detection results
├── grocery_yolov5s.pt         # Grocery-specific model (optional)
├── yolov8n.pt                 # Default YOLO model
├── yolov8n.onnx              # ONNX export (optional)
├── yolov8n_openvino_model/   # OpenVINO export (optional)
├── requirements.txt           # Python dependencies
└── readme.md                  # This file

🛠️ Installation

Prerequisites

• Python 3.8 or higher
• pip (Python package manager)
• Virtual environment (recommended)

Setup Steps

1. Clone the repository

git clone https://github.com/NSTKrishna/Video_model.git
cd Video_model

2. Create and activate virtual environment

python3 -m venv venv
source venv/bin/activate  # On macOS/Linux
# OR
venv\Scripts\activate  # On Windows

3. Install dependencies

pip install -r requirements.txt

4. Add your model (Optional - for grocery detection)
   • Place grocery_yolov5s.pt in the project root directory
   • If not available, the system will automatically use YOLOv8n

🎯 Usage

Basic Usage

Run the inventory detection on your video:

python src/run.py

Expected Output

📦 Final Inventory:
{
    "frame_based_count": {
        "mouse": 1,
        "book": 2
    },
    "tracking_based_count": {
        "mouse": 1,
        "book": 4
    }
}

Using Custom Video

Place your video file in the videos/ directory and update src/run.py:

VIDEO_PATH = "videos/your_video.mp4"

🔧 Configuration

Model Selection

The system automatically selects the best available model:

1. grocery_yolov5s.pt (if present) - Grocery-specific detection
2. yolov8n.pt (fallback) - General object detection

Adjusting Detection Settings

Edit src/detect_inventory.py to customize:

# Change image size (smaller = faster, larger = more accurate)
results = model(frame, imgsz=320, verbose=False)

# Adjust frame extraction rate in extract_frames.py
frames = extract_frames(VIDEO_PATH, fps_interval=1)  # Extract 1 frame per second

🧪 Advanced Features

Export Models for Better Performance

Run the export script to create optimized model formats:

python export_model.py

This creates:

• yolov8n.onnx - ONNX format (portable, faster)
• yolov8n_openvino_model/ - OpenVINO format (fastest on Intel CPUs)

Benchmark Performance

Test different model formats on your system:

python benchmark.py

📋 Requirements

Main dependencies:

• ultralytics - YOLO implementation
• opencv-python - Video processing
• numpy - Numerical operations
• onnxruntime - ONNX inference (optional)
• openvino-dev - OpenVINO optimization (optional)

See requirements.txt for complete list.

🎓 How It Works

1. Frame-Based Detection

• Extracts frames from video at specified intervals
• Runs YOLO detection on each frame
• Counts all detected objects across frames

2. Tracking-Based Detection

• Processes video stream continuously
• Assigns unique IDs to objects using ByteTrack
• Counts only unique objects (avoids duplicates)

3. Optimization Pipeline

Video Input → Frame Extraction → YOLO Detection (320px) → Object Counting
                                      ↓
                            Tracking (ByteTrack) → Unique ID Assignment

🐛 Troubleshooting

Issue: "grocery_yolov5s.pt not found" warning

Solution: This is normal if you don't have the grocery model. The system will use YOLOv8n as fallback. To remove the warning, either:

• Add the grocery_yolov5s.pt file to the project root, OR
• The code will continue working with YOLOv8n

Issue: Slow processing

Solutions:

1. Ensure imgsz=320 is set in detection/tracking calls
2. Use YOLOv8n instead of larger models (YOLOv8x)
3. Extract fewer frames (increase fps_interval)
4. Run python export_model.py and use ONNX/OpenVINO

Issue: CUDA/GPU errors

Solution: The code is optimized for CPU inference. GPU support is automatic if PyTorch detects CUDA.

📈 Performance Tips

1. For Speed: Use imgsz=320, YOLOv8n model, reduce frame extraction rate
2. For Accuracy: Use imgsz=640, grocery_yolov5s.pt or larger models, extract more frames
3. For Balance: Current default settings (320px, YOLOv8n, 1 fps)

📚 Documentation

Additional documentation files:

• FINAL_RESULTS.md - Complete optimization results
• GROCERY_MODEL_SETUP.md - Grocery model setup guide
• OPTIMIZATIONS.md - Detailed optimization explanations
• QUICK_REFERENCE.md - Quick command reference

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

📄 License

This project is open source and available under the MIT License.

👥 Author

NSTKrishna

• GitHub: @NSTKrishna

🙏 Acknowledgments

• Ultralytics for YOLO implementation
• OpenVINO for CPU optimization
• ONNX Runtime for cross-platform inference

📞 Support

For issues and questions:

1. Check the troubleshooting section above
2. Review documentation in the project
3. Open an issue on GitHub

---

Last Updated: November 25, 2025

Version: 2.0 (Optimized)

⭐ If you find this project useful, please give it a star!