Colloquium: Diogo Martins (C&O)
11 oktober 2017 14:30 - Locatie: Lecture room H, Faculty of Aerospace Engineering, Kluyverweg 1, Delft.
Fusion of stereo and monocular depth estimates in a self-supervised learning context
In this article we study the case where a robot is navigating in the operational environment equipped with stereo and monocular vision on-board. The main contribution is an innovative method for fusion of depth estimates from both stereo vision and a convolutional neural network (CNN) that processes a single still image. The fusion preserves high accuracy points of the stereo map and uses the output of the CNN to supervise stereo vision in the other low-confidence regions. Furthermore, we also show that the performance of the monocular estimator in the operational environment improves if stereo vision is used as supervisory input in a self-supervised learning (SSL) fashion. The merging framework is implemented on-board of a Parrot SLAMDunk and tested in real world scenarios, providing more reliable depth maps that can be used for indoors and outdoors navigation.