Short-Term Vehicle Trajectory Prediction Using Attention Mechanism Integrated GRU Network

Predicting the motion and behavior of surrounding vehicles is an essential task for motion planning and decision-making of autonomous vehicles in complex traffic conditions. In this paper, we propose a short-term vehicle trajectory prediction framework using attention mechanism integrated GRU network. We use an encoder-decoder model as the main architecture. A gate recurrent unit (GRU) coupled with temporal attention and graph attention is used to extract and fuse more important information which could be used for trajectory prediction. The temporal attention could extract temporal information and graph attention could consider interactions between surrounding vehicles within sensing range. The extracted information is fed into fully connected layers to obtain predicted trajectory. The publicly next generation simulation (NGSIM) I-80 and US-101 datasets are used to evaluate proposed model. Compared to other prediction models, our model shows improvement on final displacement error (FDE) and average displacement error (ADE). The results show that model with attention mechanism improves prediction accuracy by 1% ~5% in 5 second prediction horizon.