Child and Therapist Tracking Model

This project implements a system for tracking and classifying children and therapists in video footage using YOLOv8 object detection and custom classification models.

Overview

The system uses two YOLO models:

1. A YOLOv8s model for person detection and tracking
2. A custom YOLOv8 model for classifying detected persons as either children or therapists

The main script processes video input, detects and tracks persons, and classifies each tracked individual over time.

Prerequisites

• Python 3.x
• OpenCV (cv2)
• NumPy
• Ultralytics YOLO

Installation

1. Clone this repository
2. Install the required packages:

   pip install opencv-python numpy ultralytics
   

3. Download the necessary YOLO models:
   • YOLOv8s model for person detection
   • Custom YOLOv8 model for child/therapist classification

Usage

1. Place your input video in the project directory and name it video.mp4.
2. Run the main script:

   python inference.py
   

3. The script will process the video and output:
   • A real-time display of the annotated video
   • An output video file named predicted_output.mp4

Test Video

You can download the test video used for this project from the following link: Test Video Link

Predicted Output Video

The predicted output video, showing the tracking and classification results, can be viewed here: Predicted Output Video Link

Model Training

The custom classification model was trained using YOLOv8 and a dataset from Roboflow. For more details on the training process, refer to training.py.

Key Features

• Person detection and tracking using YOLOv8 and ByteTrack
• Custom classification of tracked individuals as children or therapists
• Temporal smoothing of classifications to improve accuracy
• Real-time visualization of tracking and classification results
• Output video generation with annotations

Configuration

You can adjust the following parameters in the script:

• time_threshold: Time interval for updating classifications (default: 3 seconds)
• confidence_threshold: Confidence threshold for high accuracy classifications
• outperform_threshold: Number of intervals a class needs to outperform the current best class

Output

The script generates an annotated video showing:

• Bounding boxes around detected persons
• Track IDs for each person
• Current classification (child or therapist) for each tracked individual

Limitations

• The system assumes that the input video contains only children and therapists.
• Classification accuracy may vary depending on the quality of the input video and the training data used for the custom model.

Future Improvements

• Implement multi-class classification for more detailed analysis
• Optimize performance for real-time processing of high-resolution videos
• Develop a user interface for easier configuration and result analysis

License

[Specify your license here]

Acknowledgements

• Ultralytics for the YOLOv8 implementation
• Roboflow for providing the dataset used in training the custom model