The goal of this project was to implement a star tracker using simulated images generated from Stellarium. The image dataset is available on Kaggle. Images were generated by incrementing right ascension and declination angles by 5 degrees in Stellarium. The images were labelled with their ra and dec.
Two methods are investigated which rely on training a supervised machine learning model to predict the sky direction in classes representing the North/South-East/West skies. These classes were derived from the right ascension and declination angles.
The first method in svm.py uses manual feature extraction to train an SVM model. Features are extracted from the brightest star (largest contour) and distance to stars in a region localized around the brightest star.
The second method in cnn.py and cnn_aug.py uses a CNN to skip the feature extraction stage and attempt to classify the images directly.