Real-time 3D stereo reconstruction for situational awareness

3D reconstruction using stereopairs is still an open research topic in different communities, such as computer vision, research and remote sensing. The handcrafted algorithms have been flanked by a growing number of deep learning based methods in the last decade, giving an additional boost to the performance of these solutions. Despite this big effort, all these solutions do not deliver 100% reliable results and, more than this, take several seconds or even minutes to deliver dense 3D reconstruction from a single stereo pair. The recent introduction of Foundation Models to train deep learning 3D reconstruction has improved their efficiency (and capacity to generalize their results) in many applications but very few algorithms aim to produce reliable dense reconstruction in real-time, despite their importance for autonomous robotic systems.

This MSc project aims to explore new solutions to deliver accurate and reliable stereo reconstructions with particular attention to real-time approaches. To reach this goal, the existing deep learning solutions available in the literature will be tested as a starting point for this work. The performance of the developed algorithm will be assessed using both terrestrial and airborne data to validate its use for terrestrial and aerial robots. At the end of this process, the student will be asked to implement some customization and modifications to the algorithm with the aim of making it more suitable for aerial robotics applications.