A Personalized Multi-Attention-Based LSTM Algorithm for Vehicle-Trajectory Prediction

Autonomous vehicles (AVs) have recently been running on the road operated by robotgo company. However, to avoid the potential risk triggered by the failure of conflict arbitration in AVs, their trajectories are essential to predict for ensuring the safety of the traffic system. This article has proposed a driving-trajectory prediction method based on a personalized multi-attention long short-term memory (PMLSTM) neural network, which attempts to reflect the natural relevance of the vehicles and surrounding environments. Specifically, a KNNbased driver-style recognition algorithm has been designed in our method to identify individualized driving styles. On this basis, LSTM networks that incorporate spatial and temporal attention mechanisms are designed for different driving styles. This network optimizes the LSTM network structure according to different driving styles and road environment information, thereby reflecting personalized requirements. Finally, extensive experiments are conducted on the NGSIM database to verify the superiority of the proposed algorithm. It is shown that our model outperforms the 4 previous approaches LSTM, S-LSTM, S-GAN, PIP and has achieved lowest root mean square error (RMSE) values across all 5 second prediction duration horizon for trajectory prediction, whose error reductions ranging from 1.82% to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{2 6. 0 3 \%}$</tex> for <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1}$</tex> second duration and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1. 9 8 \%}$</tex> to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{3 6. 2 3 \%}$</tex> for <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{5}$</tex> second duration. The proposed new paradigm constructs a new way for significant advances in trajectory prediction, especially in reducing the prediction error, taking us one step forward in realizing the fully autonomous transportation system.