Attention-Based Bidirectional LSTM Network for Target Tracking

Abstract

One of the major challenges in target tracking is the inaccurate state estimation caused by motion uncertainty of the target. In this paper, a deep neural network-based method is proposed to fit the nonlinear mapping relationship between the filtering state, the measurement, and the real state to achieve effective tracking of the maneuvering trajectory. First, the encoder method is used to mine the favorable information in the target state and radar observation. Next, the bidirectional long short-term memory network (BI-LSTM) is used to memorize and fit the potential laws of the target motion. Then, the attention mechanism is introduced to make the network automatically capture the importance of the trajectory sequence, which greatly improves the efficiency and accuracy of estimation. Finally, we use decoder method to map the complicated functions between the estimated trajectory and the real trajectory. The simulation results verify that our attention-based bidirectional LSTM (ATBI-LSTM) method has a higher estimation accuracy and better dynamic performance than traditional algorithms without prior knowledge and a complex parameter adjustment process.