All unmanned aerial vehicles that use synthetic aperture radar (SAR) systems are equipped with inertial navigation systems (INS) to reduce motion error. Additional motion compensation (MOCOMP) from the data itself is still necessary to achieve required accuracy of a SAR. An affordable method for small drones has yet to be created. We propose machine learning with deep convolutional neural network (CNN) to extract motion error such as sway (right and left) and surge (forward). Results show that the CNN is capable of recognizing gradual drone motion deviations. It has the potential to pick up sudden motion error as well, overcoming major deficiencies of traditional MOCOMP methods, and the need for INS.
MEMC-Net: Motion Estimation and Motion Compensation Driven Neural Network for Video Interpolation and Enhancement
