GYM-Net: Graphical Y-branched Multi-tasking Network for Trajectory Prediction

Abstract

For autonomous driving, the ability to predict the future from the past trajectory is crucial. In this study, a graphical Y-branched multi-tasking network (GYM-Net) is proposed. The GYM-Net predicts future trajectory from multi-scale graphical features using multi-task learning. The GYM-Net predicts both the final position after ten-time steps and the future trajectory after ten-time steps based on the previous trajectory of five-time steps. Using the NBA dataset that is widely used to predict the agents that are active in multi-agent trajectory predictions, we assessed the performance of our model. The performance of our model is 2.37 m, 2.31 m, 3.87 m, and 3.81 m for the evaluation metrics of avgADE, minADE, avgFDE, and minFDE, respectively. Compared with the performance of the graph U-Net, our model shows a decrease in error by 0.02, 0.04, 0.02, and 0.04, respectively. Consequently, our model can predict the trajectory of multiple dynamic agents.