Robust trajectory forecasting in autonomous systems using mixtures of Student's T-distributions with T-DistNet

This paper addresses the challenge of predicting future trajectories of agents in complex traffic scenes, emphasizing the need for reliable predictions that are robust to various sources of uncertainty. Current methods for trajectory prediction often overlook the uncertainty aspect, although typically relying on deep neural networks (DNNs) trained to predict mixtures of Gaussian and Laplace distributions. In our study, we evaluate the significance of distribution choice for achieving reliable and robust predictions in uncertain environments and introduce T-DistNet, which employs a mixture of Student's T-distributions for superior uncertainty modeling. This approach enables more accurate performance in scenarios with varying levels of uncertainty compared to other mixed distributions. Our analysis demonstrates that T-DistNet effectively models uncertainty, facilitating efficient and precise predictions.