Attention based graph convolutional networks for trajectory prediction

Predicting the future trajectory of different traffic agents in the complex traffic environments plays an important role in keeping the driving safety of self-driving cars, especially on urban roads. In the most of the existing works, researchers always use the long short-term memory network (LSTM) to solve this problem, since the LSTM has powerful capability for capturing the temporal dependency in motion trajectory. However, they only consider forward time cues and ignore the spatial-temporal correlations between traffic agents. Inspired by the previous work which utilizing the spatial-temporal graph, we design a spatial attention based spatial-temporal graph convolutional network, which assigns different attention weight to the the graph to take the different social interactions among the self-driving cars into consideration. We conduct extensive experiments on the benchmark InD to compare our method against many baselines. The experiment results indicate the superiority of our method than previous method, about 22% and 17% improvement on the metric of ADE and FDE respectively.