SURF

Zisu Dong

3D City-scape Reconstruction from Motion

Reconstructing dense models of real-world 3D scenes is important for autonomous driving tasks. However, motion estimation for an agile single camera moving through general, unknown scenes has proved to be a challenging problem. The task becomes much more challenging in autonomous driving when real-time performance is required under disturbance of transient change of moving objects (e.g. vehicles, pedestrians) and surrounding environments (e.g. lighting conditions). The goal of this project is to build a pipeline for processing driving videos gathered by on-dash car camera on top of existing Structure from Motion method. This pipeline should generate high-quality reconstruction model of the environment using previously unknown scene along with mapping the trajectory of the 3D camera pose as output. The reconstructed vision system should ideally enable localization, general spatial awareness and scene understanding, opens up new possibilities for learning of autonomous driving policies.

Message to Sponsor

I sincerely appreciate the Pergo Fund for providing me this opportunity to really dive into a meaningful research project. During the summer, I explored the current cutting-edge technologies of autonomous vehicle, which is at the intersection of theory and real-life applications. This experience helped me to confirm my interest in pursuing a research career, exposed me to the computer vision community and provided me high-quality talks on how to extend the impact of my research.
  • Major: Computer Science
  • Sponsor: Pergo L&S
  • Mentor: Fisher Yu