TARGET POSE ESTIMATION USING DEEP LEARNING

Abstract

This thesis aims to enhance the field of target pose estimation from 2D target images using deep learning techniques. To achieve this, a cutting-edge convolutional neural network, known as High-Resolution Net, was employed to train a keypoint detection model and assess its performance. The experiment utilized two diverse datasets comprising 600,000 synthetic images and 77,077 High Energy Laser Beam Control Research Testbed (HBCRT) images. These images are of six different unmanned aerial vehicles that were utilized for training and evaluation purposes, with High-Resolution Net being trained on 80% of the images and tested on the remaining 20%. The MMPose framework, a Python library with multiple options for convolutional neural networks, was utilized to run High-Resolution Net. The findings revealed that High-Resolution Net performs well in pose estimation, but a significant gap in left and right inversion remains due to the symmetry of the target shape. This research serves as a stepping stone for future target pose estimation studies utilizing High-Resolution Net. Further research will concentrate on improving the accuracy of left-right discrimination in libraries to enhance these outcomes.