DEVELOPMENT OF NEURAL NETWORK BASED STAR PATTERN RECOGNITION SYSTEM USING RADIAL BASED FEATURE EXTRACTION AND DEPLOYMENT ON RASPBERRY PI COMPUTER

• Main Author: Mohammed Catraguna, Brian
• Format: Final Project
• Language: Indonesia
• Subjects: Teknik (Rekayasa, enjinering dan kegiatan berkaitan)
• Online Access: https://digilib.itb.ac.id/gdl/view/57901
• Institution: Institut Teknologi Bandung

The need for a more efficient star sensors is continually increasing, providing the need for faster star pattern recognition system. The star sensor provides the most accurate sensor for attitude determination. With the rise of the space industry and more micro satellites are being produced, the need for a small sized and efficient star sensors need to be developed to suit the demands of the industry. Most complex star pattern recognition systems require the need for ground – based systems and personnel for computing power. Conventional on-board systems experience computational inefficiency due to the memory allocated for the star catalogue and the algorithms that require a brute force search on the entire star catalogue. New computation architectures based on the anatomy of the human brain promise a high accuracy and high-speed computing that surpasses the capability of conventional serial processing. The motivation of using neural network or deep learning is to replace this brute force search of the entire star catalogue that has been proven to be inefficient. A deep learning model requires feature extraction from a star image in order to reduce its dimensionality for more efficiency of data use. The radial based feature extraction is a method to extract features from a given star image in a way that it is suitable for neural networks because of its simple data structure of an array. Such a system can determine in real time, which direction a star sensor is facing with reliable speeds needed for more complex missions. However, the system needs to also reliable when operating on small computers in order to be deployed on a micro satellite. Therefore, Raspberry Pi computer with the size of only 2.22 by 3.37 inches can be used as a benchmark to test out the algorithm that will soon be running on a micro satellite. Monte carlo simulations is also done to simulate the real conditions of the space where the star sensor will be faced to a random attitude. The results of the deep learning model’s accuracy is promising with the best case of 98% when given the optimal conditions while can also have worse scenario of an accuracy below 40% when given many noises. The processing time of the system when run in a Raspberry Pi is also considered quick when it only takes around 1.4 seconds. Monte carlo simulations were also promising when using the best model can have an accuracy of up to 99%. Based on this thesis, the implementation of neural networks and deep learning can be validated to be used in a star pattern recognition system using radial based feature extraction.