Optimizing Multi-Layer LSTM Based on Non-Local Spatiotemporal Interactions for Vehicle Trajectory Prediction

Abstract

Accurate trajectory prediction is critical to improve the planning and control capabilities of autonomous vehicles. In complex traffic scenarios, the influence of social interactions among vehicles plays an important role in shaping their future trajectories. Vehicle trajectory models based on recurrent neural networks (RNNs) or convolution neural networks (CNNs) often struggle to perform well in long prediction horizons. The majority of existing approaches limit themselves to a fixed spatial neighborhood or a short temporal window for interaction modeling, which accumulate errors and overlook long-range dependencies. To overcome these limitations, a non-local spatio-temporal interaction-based optimized long-short-term memory (NST-LSTM) model is introduced to predict future trajectories. The proposed model effectively captures high- order interactions among vehicles without being constrained to a fixed spatial neighborhood or a limited temporal window. A series of experiments are carried out using real-world High-D dataset. The experimental results reveal that our model outperforms several baseline models and achieves 80% RMSE reduction relative to the next-best LS-LSTM across the five-second prediction horizon. A detailed ablation study is also conducted to choose optimal hyper-parameters for the LSTM model including depth, learning rate, and batch size.