Velocity Prediction for Multi-rotor UAVs Based on Machine Learning

Abstract

Currently, multi-rotor UAV’s navigation mainly depends on satellite navigation systems. In many environments (e.g., indoor, urban, or canyon), lack of satellite signal will lead UAV navigation to failure. In that case, one backup solution is to use the inertial navigation method. Inertial navigation method integrates measurements from gyroscope and accelerometer to obtain the orientation, speed, and position. Due to the instability of sensor bias, a slight orientation error caused by gyroscopic bias change will lead to enormous position error. Noting that there is a strong correlation between the pose and velocity for a flying multi-rotor UAV, we can search for a solution to calculate velocity directly from UAV’s pose. In this work, we propose to use support vector machine (SVM)-based machine learning technique to predict the moving speed of the aircraft. This approach builds a relationship directly between the orientation data and velocity by training. The experiment has two stages. In early stage, we have tested our method in simulation environment; then at later stage, we have tested our method in real-world cases. Experimental results our method can predict the UAV’s velocity within the error of 0.3 m/s and the squared correlation coefficient between the predicted velocity and the ground truth is about 0.8. The method can be used as a complementary navigation source to achieve higher localization accuracy and stability.