Encoder-Decoder Multi-Step Trajectory Prediction Technology Based on LSTM

Abstract

Aiming at the trajectory and motion characteristics of weakly constrained non-cooperative targets, a LSTM-based encoder-decoder multi-step trajectory prediction technology (EDMTP) is proposed. The introduction of first-order difference processing reduces the time dependence of the trajectory data, and obtains a trendless trajectory. Constructing an input and output trajectory data pair, transforming the prediction problem into a supervised learning problem, the change of model performance in the multi-step prediction process is studied to realize end-to-end trajectory prediction. Simulation results show that this method can extract more trajectory features from historical trajectory data, and has obvious advantages in multi-step trajectory prediction. Compared with the trajectory prediction algorithms of KFTP and HMMTP, the error growth rate of EDMTP decrease by 2.18% and 3.52% year-on-year, respectively, and achieves better trajectory prediction results.