Spatial Temporal Graph Fusion Network for Trajectory Prediction of Moving Targets

How to accurately predict trajectories of surrounding moving objects for the autonomous vehicle is still a challenging problem because trajectories of these objects are influenced not only by themselves, but also by their interactions with each other. Previous works based on deep learning processed spatial and temporal information separately, which cannot very well describe these interactions among the moving objects at different time intervals. In this paper, we propose a spatial temporal graph fusion network (STGFN) to capture the spatial and temporal information simultaneously. Specifically, a new 3D graph architecture is designed to incorporate both spatial and temporal edges, which is used to represent these interactions of moving objects. Then, the graph attention network (GAT) is employed to explicitly focus on these significant interactions. And at last, encoder-decoder convolutional gated recurrent units (ConvGRU) are used to carry out accurate predictions of different types of traffic agents. To evaluate STGFN performances, the trajectory dataset for urban streets ApolloScape is used. Results show that our proposed STGFN outperforms several baseline methods on both the weighted sum of average and final displacement error.