Transfollower: Long-Sequence Car-Following Trajectory Prediction Through Transformer

Car-following refers to a control process in which the following vehicle (FV) tries keep safe distance between itself and lead (LV) by adjusting its acceleration response actions of ahead. The corresponding car-following models, describe how one follows another traffic flow, form cornerstone for microscopic simulation intelligent development. One major motivation models is replicate human drivers' longitudinal driving trajectories. To model long-term dependency future on historical situations, we developed long-sequence trajectory prediction based attention-based Transformer model. general format encoder-decoder architecture. encoder takes speed spacing data as inputs forms mixed representation context using multi-head self-attention. decoder LV profile input outputs predicted FV generative way (instead an auto-regressive way, avoiding compounding errors). Through cross-attention decoder, learns build connection speed, can be obtained. We train test our with 112,597 real-world events extracted from Shanghai Naturalistic Driving Study (SH-NDS). Results show that outperforms traditional driver (IDM), fully connected neural network model, long short-term memory (LSTM) terms accuracy. also visualized self-attention heatmaps explain derives predictions.