Intelligent Vehicles Trajectory Prediction with Spatial and Temporal Attention Mechanism

Abstract

Abstract The prediction of surrounding vehicle trajectory is an important research contents related to the safety of intelligent vehicles, and the strong non-linear and randomness add to the difficulty of the prediction task. In response to this challenge, this paper uses LSTM (Long Short-Term Memory) networks as a prediction framework, considering the spatial constraint relationship of the target vehicle, and proposes a spatial attention mechanism to distinguish vehicle interactions under the influence of different spatial locations; in order to capture the context information of the target vehicle, a LSTM model with the temporal attention mechanism is proposed, and key historical trajectory information is extracted for training. The experiments are constructed on the NGSIM dataset, and the experiments confirm that our prediction framework combined with the dual attention mechanism can achieve the leading performance in the synchronization method.