Neural Network Based Tracking of Maneuvering Unmanned Aerial Vehicles
2022 56TH ASILOMAR CONFERENCE ON SIGNALS, SYSTEMS, AND COMPUTERS, pp. 380–386.
The motion model of an unmanned aerial vehicle (UAV) is a time-varying function that is often unknown to the tracking entity, which makes accurate tracking of highly maneuvering UAVs a challenging problem. Most state-of-the-art tracking techniques employ a fixed set of parametric models that approximate the possible maneuvers along the target trajectory to a reasonable accuracy. However, such predetermined motion models might not be adequate for frequent and aggressive maneuvers performed by small UAVs. To this end, we build a data driven adaptive filtering algorithm that improves the tracking accuracy by using a recurrent neural network (RNN)-based motion model that is trained on realistic simulated data generated from a medium fidelity simulink model of a fixed-wing UAV. We then train another feed-forward neural network in conjunction with the pretrained RNN-based motion model, to adaptively combine the incoming current measurement with the predicted state based on the output of a change detection algorithm that detects any increase/decrease in the uncertainty in the predicted states. Our analysis and results show that the proposed tracking algorithm outperforms the state-of-the-art interactive multiple model (IMM) algorithm for highly maneuvering trajectories.