Feature Matching and Deep Learning Models for Attitude Estimation on a Micro-Aerial Vehicle

In today’s digital era, destructive and non-destructive methods of cyber-attacks are being exploited particularly for robotic applications. For this, artificial intelligence particularly in cybernetics plays a vital role. In this study, we present different neural network models to address such issues based on deep learning approaches. We further apply a feature matching algorithm on the publicly available EuRoC dataset to identify a persistent feature in continuous frames. The flight path of Micro Aerial Vehicles (MAVs) was estimated using the Inertial Measurement Unit (IMU)’s, gyroscope, and accelerometer data recorded at 200 Hz. An adaptive filter was used for the prediction of quaternions at each time step for comparison. We implemented two deep learning models, Convolutional Neural network (CNN) and Recurrent Neural Network (RNN), as an alternative to the adaptive Madgwick’s algorithm and for estimation of spatial orientation of the MAV at every time step and computed the results. We also performed benchmarking for the RNN model by comparing it to the existing results observed using Madgwick’s algorithm.