Research on the improvement of the LaneGCN trajectory prediction algorithm

The LaneGCN proposed by Uber has achieved good performance in trajectory prediction, but it has shortcomings in capturing long range information, expressing road information and modelling the strong and weak relationships of interaction between actors. In this paper, the LaneGCN is improved from three parts. Firstly, multi-scale long short-term memory is introduced to encode multi-scale trajectory information. Secondly, relative distance information is added to enhance the spatial expressive capacity of the model in the process of road information encoding. Finally, we build a weighted interaction model based on Graph Attention Networks in the process of road information encoding. In order to verify the performance of the improved model, ablation and comparison experiments are designed in this paper. The results showed that all the evaluation metrics are lower than the LaneGCN and the overall performance of the model is improved.