Stable Learning-based Interest Points for Real-time SLAM

Nowadays, Visual SLAM (Visual Simultaneously Localization and Mapping, VSLAM) is an essential part of many applications, such as mobile robots, augmented reality, virtual reality, and autonomous driving. Although VSLAM has experienced rapid development, state-of-the-art methods are still facing the challenge of achieving accuracy, robustness, and real-time performance simultaneously. In this research, it is hoped that the advantages of deep learning in 2D images can be combined with a complete VSLAM framework, so as to improve the robustness and accuracy of the overall system while satisfying real-time performance. Specifically,I plan to use a deep learning-based method to extract interest points on the front end to ensure the robustness and accuracy of points extraction. At the same time, more diverse methods for location recognition will be used so that the SLAM system could be more robust.

Promotors: Ming Cao and Jacquelien Scherpen
PhD: Qihang Zhang