Classification of Aerial Platform Using Deep Learning

Abstract

UAVs have made significant progress in recent years, with several corporations exploring their usage for delivery. It is expected that more such objects would fly over the city in near future. The ground detection of flying objects, will enable UAVs navigate and adjust to other objects in their path. Flying objects must be recognized from the ground for UAV route monitoring and guidance, taking into account other flying objects in the path. This thesis aims to investigate the use of machine learning to the problem of aerial platform detection. The study includes the complete process of applying and assessing machine learning methods to subject issue. The selected AI model for this study was YOLOv5. To assess the model’s performance and robustness to the given problem, different feature extraction algorithms were used to reconstruct images from features before training the YOLOv5 network on these images. Feature extraction algorithms were applied at each channel of RGB image and then reconstructed images from features were concatenated to form a single tri channel image. The results show that features extracted from Radon transform, Fan-beam transform and Discrete Cosine transform combined together to construct a tri channel image yields better prediction of the classes then other algorithms.