Canche, Mario, Ranganath Krishnan, and Jean-Bernard Hayet. "Epistemic Uncertainty Quantification in Human Trajectory Prediction." ICT Applications for Smart Cities. Springer, Cham, 2022. 37-56.
Human Trajectory Prediction (HTP) is a critical technology in several areas related to the development of smart cities, such as video monitoring or autonomous driving. In the last years, there has been a leap forward in the state of the art of HTP, with great improvements observed in most of the classical benchmarks used in the related literature. This has been possible through the use of powerful data-driven deep learning techniques, coupled with probabilistic generative models and methodologies to cope with contextual information and social interactions between agents. In this chapter, we first show how incorporating Bayesian Deep Learning (BDL) techniques in Human Trajectory Prediction allows to provide realistic estimates of the epistemic and aleatoric uncertainties on the computed predictions. In addition, we also present an original methodology to assess the quality of the produced uncertainties (through BDL or other probabilistic approach).