Simultaneous Localization and Mapping (SLAM) denotes the technique of ego-motion tracking and constructing or updating a map of unknown surroundings at the same time. It plays a central role in a variety of application scenarios, such as autonomous driving, virtual/augmented reality, etc. In order to endow better autonomy and interaction capability for mobile robots, techniques such as vision-based semantic understanding are required and have become an active research field in recent years. Though the deep learning-based methods have shown impressive results for image processing such as semantic segmentation, there is still much potential to exploit for making robotic vision more robust, efficient and semantically accurate. In this project, novel non-parametric
learning methods will be investigated and developed for robotic perception in the context of SLAM. In particular, we focus on learning methods proposed on the Riemannian manifolds such that better tracking and mapping performance are expected both geometrically and semantically. More specifically, the project is composed of the
following work packages:
- Literature review of existing vision-based semantic SLAM frameworks and non-parametric learning based on Riemannian geometry, especially techniques from geometric deep learning and Gaussian process.
- Development of a proper vision-based 3D mapping method, based on which a learning framework for semantic understanding is then proposed.
- Development of a SLAM framework and deployment on a real robotic platform mounted with visual sensors, e.g., LiDAR and stereo-camera.
- Evaluation based on real-world dataset and experiments in multiple scenarios, e.g., 3D object detection, large-scale tracking and semantic reconstruction, etc