Context-free Vehicle Trajectory Prediction

The following thesis aims at analysing the usage of Cooperative Awareness Messages (CAM) for predicting future trajectories of vehicles in a context-free environment. An initial focus has been devoted for creating and training a neural network, namely Graph-based Trajectory Prediction (GraphTP), using one of the most famous dataset available for vehicle trajectory prediction: Argoverse 2 Motion Forecasting Dataset. This initial step has been fundamental to compare the accuracy of the network to other competitors using a well-known dataset. Simultaneously, CAM data extracted from Road-Side Units present in Modena have been processed to create a second custom dataset with the same structure as Argoverse. Lastly, both GraphTP and the competitors have been evaluated on such dataset in two situations: zero-shot and finetuning.