Modeling Vehicle Interactions via Modified LSTM Models for Trajectory Prediction

Abstract

The long short-term memory (LSTM) model is one of the most commonly used vehicle trajectory predicting models. In this paper, we study two problems of the existing LSTM models for long-term trajectory prediction in dense traffic. First, the existing LSTM models cannot simultaneously describe the spatial interactions between different vehicles and the temporal relations between the trajectory time series. Thus, the existing models cannot accurately estimate the influence of the interactions in dense traffic. Second, the basic LSTM models often suffer from vanishing gradient problem and are, thus, hard to train for long time series. These two problems sometimes lead to large prediction errors in vehicle trajectory predicting. In this paper, we proposed a spatio-temporal LSTM-based trajectory prediction model (ST-LSTM) which includes two modifications. We embed spatial interactions into LSTM models to implicitly measure the interactions between neighboring vehicles. We also introduce shortcut connections between the inputs and the outputs of two consecutive LSTM layers to handle gradient vanishment. The proposed new model is evaluated on the I-80 and US-101 datasets. Results show that our new model has a higher trajectory predicting accuracy than one state-of-the-art model [maneuver-LSTM (M-LSTM)].

Keywords

• Trajectory prediction
• vehicle interactions
• shortcut connection
• long short-term memory (LSTM)