Project Description: Tracking Space Debris and Other Objects Through Camera
Overview This project involves tracking space debris and other objects through a camera using a custom-trained YOLOv8 model. The goal was to detect and track multiple classes of objects in images & video, ensuring high accuracy and robustness of the model.
Dataset Preparation
Initial Dataset: Utilized an existing dataset but chose not to use its labels due to their subpar quality. Link- https://drive.google.com/drive/folders/1cYUHXstqUCjXKG7hKuKqtUsWeQQ_3cYz?usp=sharing
Manual Labeling: Manually boxed and labeled objects in the dataset. Selected 120 images for each of the 11 classes, resulting in a total of 1,320 manually labeled images.
Data Augmentation: Applied horizontal and vertical flips, and resized all images to 640x640 pixels, expanding the dataset to 2467 images.
Data Split: Divided the dataset into training (70%), validation (20%), and test (10%) sets using Roboflow, ensuring a comprehensive and balanced dataset for model training.
Dataset foramt: Organiszing the images and labels as per YOLOv8 provided format for training and generating the yaml file.
Final Dataset: 2467 fined labelled, resized, randomly flipped dataset on yolov8 format. Link- https://www.kaggle.com/datasets/muhammadzakria2001/space-debris-detection-dataset-for-yolov8
Environment Setup
YOLOv8 Loading: Leveraged the Ultralytics YOLOv8 library for object detection.
Kaggle Environment: Utilized Kaggle’s resources for model training, including the latest CUDA, two Tesla T4 GPUs, an Intel Xeon 4-core CPU, 30 GB RAM, and 100 GB disk space.
Model Architecture
Model Training
Configuration:
Epochs: 100
Image Size: 640x640 pixels RGB
Optimizer: Adam
Training Duration: Completed the custom training within approximately one hour.
Model Testing
Performance: Evaluated the model using the test set, achieving highly accurate and satisfactory results.
Confusion matrix: 
Training statistics: 
Plots:
Labels
Prediction
Conclusion This project demonstrates a comprehensive approach to preparing a high-quality dataset, setting up an efficient training environment, and successfully training and testing a YOLOv8 model for tracking space debris and other objects with impressive accuracy.