Multi-Step Ballistic Vehicle Trajectory Forecasting Using Deep Learning Models

Abstract

This paper compares several deep learning models on the task of multi-step trajectory forecasting of a non-manoeuvring ballistic vehicle. We use state-of-the-art techniques to build and train LSTM, GRU, and Transformer architectures and test their performance versus the multi-layer perceptron baseline. The experiments on synthetic data show that, in our problem settings, trajectory forecasting is best performed with the LSTM network with a trainable initial state. Although the Transformer models were able to outperform the baseline, they could not outrun the recursive neural networks in terms of prediction errors.