Position: PhD Candidate
Current Institution: University of Pennsylvania
Abstract: Using deep learning for state estimation in a Bayesian framework
The advancements of deep learning approaches give rise to new possibilities in various fields of studies. One particular area that draws such research attention is visual inertial navigation systems (VINS). After decades of development these systems achieve a performance of superior accuracy robustness and efficiency with the lightweight and low-cost sensory suite of cameras and IMUs. Although highly developed further improvement with deep learning is still of keen interest to the community. While some are interested in developing end-to-end network structures that redefine the current framework applying learning to more specific parts of the system such as image feature descriptor occlusion and depth prediction also shows promising results. More interestingly recent work shows that deep learning can effectively identify patterns in IMU data for pedestrian applications overcoming the limit of its mathematical motion model and achieves low-drift long-term tracking without the aid of external sensors such as cameras. The theme of my research has been incorporating the advantage of deep learning approaches into the existing state estimation frameworks. I investigate VINS algorithms on resource constrained platforms and more recently I have been experimenting with inertial navigation systems (INS) to develop ways to effectively utilize deep learning in a Bayesian framework and to expand the often neglected power of the noisy internal sensors.
Wenxin Liu is a PhD student at University of Pennsylvania working with Vijay Kumar and Kostas Daniilidis. Her research interest focuses on state estimation with deep learning on resource constrained platforms.