A return-to-home unmanned aerial vehicle navigation solution in global positioning system denied environments via bidirectional long short-term memory reverse flightpath prediction

In this paper, bidirectional long short-term memory (B-LSTM) deep learning modeling is proposed as an approach to facilitate autonomous return-to-home (RTH) aerial navigation in environments with compromised global positioning system (GPS) reception. Logged samples of ten radiometric features are extracted from onboard sensors (i.e., accelerometer, barometer, GPS, gyroscope, magnetometer) in two outdoor experimental scenarios of different altitudes and velocities. These samples are used for training and validating B-LSTM models with single and parallel architectures. The former architecture consists of a single B-LSTM model that processes all input features across the x-, y-, and z-axes to predict a three-dimensional local position, whereas the latter comprises three parallel B-LSTM models, each for processing only the features of a specific dimension (i.e., x, y, or z) and predicting local position in the respective axis. Evaluations demonstrate the validity of the proposed approach, with a 4-m average mean square error (MSE). This is achieved without imposing resource-consuming computational overhead, modifications to existing hardware, or changes to physical infrastructure and communication protocols. Due to using existing onboard sensors and accommodating varied scenarios, the proposed approach finds applications in autonomous navigation, including unmanned aerial vehicles (UAVs) and ground vehicles.