Dynamic Scene Understanding

Lead Research Organisation: University College London
Department Name: Computer Science

Abstract

This thesis will focus on the problem of finding compact yet dense representations of scene geometry for real-time 3D perception systems. So far, SLAM systems for robotics applications represent the scene using fully dense or sparse scene points. While dense 3D maps are attractive for capturing surface details and being easily augmented with semantic scene information they are extremely costly to store and process. Sparse feature-based maps avoid these computational costs, but instead capture only partial scene geometry and cannot be used for robotic manipulation tasks such as grasping.
This thesis will focus on discovering new compact representations of scene geometry that are dense and can capture scene semantics and using them within real-time SLAM systems for dynamic scene reconstruction. The focus will be on representing scenes at the level of objects with compact low dimensional code-based encodings and to use this high level representation for data association, tracking and mapping of complex dynamic scenes with the final goal of making them available for robotic tasks

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/S021566/1 01/04/2019 30/09/2027
2251565 Studentship EP/S021566/1 23/09/2019 22/09/2023 Jingwen Wang