Skip to content Skip to navigation
Qualifying Exam
5/9/2019 10:30 am
CoRE A 301

A new approach for prediction of pedestrian trajectories based only on images.

Blerta Lindqvist, Rutgers University

Examination Committee: Prof. Kristin Dana, Prof. Casimir Kulikowski, Prof. Kostas Bekris and Prof. Desheng Zhang

Abstract

Proliferating autonomous devices necessitate prediction of pedestrian trajectories in order to navigate realistic environments. Such prediction is currently done using a non-holistic approach based on pedestrian coordinates, with additional intertwined information from scene images. This paper presents a holistic approach that generates image predictions based on image inputs to facilitate more efficient decision making in autonomous devices. This is made possible by manipulating the common architecture from a GAN with LSTM encoder and decoder to a GAN where the LSTM sits in the low dimensionality space of the GAN.