AG 1, INET, AG 5, RG1, SWS, AG 2, AG 4, D6, AG 3   Info

Public Audience

English

60 Minutes

Saarbrücken

Humans have an innate ability to excel at activities that involve prediction of complex object dynamics such as predicting the possible trajectory of a billiard ball after it has been hit by the player or the prediction of motion of pedestrians while on the road. A key feature that enables humans to perform such tasks is anticipation. To advance the field of autonomous systems, particularly, self-driving agents, in this thesis, we focus on the task of future prediction in diverse real world settings, ranging from deterministic scenarios such as prediction of paths of balls on a billiard table to the predicting the future of non-deterministic street scenes. Specifically, we identify certain core challenges for long-term future prediction: long-term prediction, uncertainty, multi-modality, and exact inference. To address these challenges, this thesis makes the following core contributions. Firstly, for accurate long-term predictions, we develop approaches that effectively utilize available observed information in the form of image boundaries in videos or interactions in street scenes. Secondly, as uncertainty increases into the future in case of non-deterministic scenarios, we leverage Bayesian inference frameworks to capture calibrated distributions of likely future events. Finally, to further improve performance in highly-multimodal non-deterministic scenarios such as street scenes, we develop deep generative models based on conditional variational autoencoders as well as normalizing flow based exact inference methods. Furthermore, we introduce a novel dataset with dense pedestrian-vehicle interactions to further aid the development of anticipation methods for autonomous driving applications in urban environments.

Connie Balzert

+49 681 9325 2000

Zoom

public