Graph Neural Network with RNNs based trajectory prediction of dynamic agents for autonomous vehicle

Trajectory prediction is an essential ability for the intelligent transportation system to navigate through complex traffic scenes. In recent times, trajectory prediction has become an important task, especially in crowded scenes, because of the great demands of emerging artificial intelligence applications like service bots and autonomous cars. As autonomous vehicles travel in interactive and highly uncertain environments shared with other dynamic road agents like other vehicles or pedestrians, predicting the trajectories of the surrounding agents is essential for an autonomous driving system (ADS) to plan safe motion, fast reaction time and comfortable maneuvers. The trajectory for each dynamic object (or road agent) is described as a sequence of states within a time interval, with each state representing the object's spatial coordinates under the world coordinate frame. In the trajectory prediction (TP) problem, given the trajectory of each object between intervals of time, we predict their trajectories between these intervals of time. We plan to design a Multi-Scale Graph Neural Network (GNN) with temporal features architecture for this prediction problem. Experiments show that our model effectively captures comprehensive Spatio-temporal correlations through modeling GNN with temporal features for TP and consistently surpasses the existing state-of-the-art methods on three real-world datasets for trajectory. Compared to prior methods, our model's performance is more for the sparse datasets than for the dense datasets.