Vehicle trajectory prediction model based on transformer and GCN
Yan Zhang, Xiaoyu Han, Wenqiang Chen · 2025
Trajectory prediction is crucial for self-driving vehicles, and accurate trajectory prediction can improve the safety and efficiency of transportation. In this paper, we presented a spatio-temporal feature fusion vehicle trajectory prediction model based on Transformer and Graph Convolutional Network (GCN),and the model mainly contains a time series coding module, a vehicle interaction coding module and a spatio-temporal feature fusion module. The time series encoding module is constructed based on Transformer, which takes the historical trajectory of each vehicle as input and outputs a set of highdimensional vectors with more time-dependent relationships; the vehicle interaction encoding module is constructed based on GCN, compared with the traditional neural network, it can consider vehicles as nodes and use directed graphs to characterize the vehicle interaction , and the output is the vehicle interaction feature vector; and the feature fusion module is constructed based on Transformer, which takes the target vehicle's high-dimensional vectors and vehicle interaction feature vectors as inputs, the output vectors are fused with spatio-temporal interaction features, and finally the future trajectories are predicted based on the fully connected layer decoder.We evaluate our model using the publicly available NGSIM US-101 and I-80 datasets. Our results show that Transformer and GCN have significant advantages in trajectory prediction compared with traditional models.