Prediction of Autonomous Vehicle Trajectories in Turnaround Scenarios

Trajectory prediction of road agents plays important role in road traffic safety. Both autonomous vehicles and roadside units can benefit from it. Other than the prediction accuracy, many design constraints should be taken into consideration like the simplicity of the model. A simple model will be executed using the limited available resources of IoT units in real-time. In this paper, a model based on a Graph Neural Network is proposed that uses both the spatial and temporal information for trajectory prediction in a fully context-aware environment. The model solves the normal issues of Deep Graph Neural Networks like over-smoothing. It also predicts the trajectory of all the surrounding agents collectively without the need to iterate it. Also, the computational workload can be distributed among road agents using their communication capabilities to cooperate in predicting the trajectories.