Person Detection using YOLOv8n (ONNX)

This repository contains a lightweight, efficient script for detecting persons in images using the YOLOv8 Nano model in ONNX format. It is designed to be simple to understand and easy to run/deploy without needing the full Ultralytics YOLO package.

📂 Project Structure

person_detection_using_yolov8n_onnx/
├── models/
│   └── yolov8n.onnx      # Pre-trained YOLOv8 Nano model
├── detect.py             # Main detection script
├── test.jpg              # Sample input image
├── README.md             # This documentation
└── ...

🛠️ Prerequisites

To run this code, you need Python installed along with the following libraries:

• OpenCV (cv2): For image processing and visualization.
• NumPy: For numerical operations and array manipulation.
• ONNX Runtime: For loading and running the ONNX model efficiently.

🚀 Installation

You can install the required dependencies using pip:

pip install opencv-python numpy onnxruntime

(Note: If you have a GPU, you might want to install onnxruntime-gpu instead for faster inference).

🏃 Usage

1. Place your image: Ensure you have an image named test.jpg in the project root directory (or update the IMAGE_PATH variable in detect.py to point to your image).

2. Run the script:

   python detect.py

3. View Results: The script will open a window named "Result" showing the detected persons with bounding boxes and confidence scores. It will also print the total count of detected persons in the console.

🧠 Code Explanation

This script (detect.py) is capable of:

1. Loading the Model: It initializes an ONNX Inference Session with models/yolov8n.onnx.
2. Preprocessing:
   • The input image is resized to 640x640 (the standard input size for YOLOv8).
   • It is converted from BGR to RGB.
   • Pixel values are normalized to [0, 1].
   • Approximately reshapes the dimensions to (Batch, Channels, Height, Width).
3. Inference: The processed image is passed through the model to get raw predictions.
4. Post-Processing:
   • Filtering: We iterate through the predictions and only keep those that match the Person class (Class ID 0) and have a confidence score higher than 0.5.
   • Bounding Box Scaling: The coordinates are scaled back from the 640x640 model size to the original image dimensions.
   • NMS (Non-Maximum Suppression): We use cv2.dnn.NMSBoxes to remove overlapping boxes and keep only the best detection for each person.
5. Visualization: Finally, it draws green bounding boxes and labels on the image and displays the result.

⚙️ Configuration

You can easily tweak these variables at the top of detect.py:

• CONF_THRESHOLD: Minimum confidence score to detect a person (default: 0.5).
• NMS_THRESHOLD: Threshold for removing overlapping boxes (default: 0.45).
• IMAGE_PATH: Path to the input image.

---

Created for review and educational purposes.