Shivaswamy, Nikhitha, Kwasnitschka, Tom and Köser, Kevin (2021) Learning Visual Free Space Detection for Deep-Diving Robots. In: Pattern Recognition. ICPR International Workshops and Challenges. , ed. by Del Bimbo, Alberto, Cucchiara, Rita, Sclaroff, Stan, Farinella, Giovanni Maria, Mei, Tao, Bertini, Marco, Escalante, Hugo Jair and Vezzani, Roberto. Springer, Cham, pp. 398-413. DOI 10.1007/978-3-030-68790-8_31.

Text Shivaswamy.pdf - Reprinted Version Restricted to Registered users only Download (4MB)

Abstract

Since the sunlight only penetrates a few hundred meters into the ocean, deep-diving robots have to bring their own light sources for imaging the deep sea, e.g., to inspect hydrothermal vent fields. Such co-moving light sources mounted not very far from a camera introduce uneven illumination and dynamic patterns on seafloor structures but also illuminate particles in the water column and create scattered light in the illuminated volume in front of the camera. In this scenario, a key challenge for forward-looking robots inspecting vertical structures in complex terrain is to identify free space (water) for navigation. At the same time, visual SLAM and 3D reconstruction algorithms should only map rigid structures, but not get distracted by apparent patterns in the water, which often resulted in very noisy maps or 3D models with many artefacts. Both challenges, free space detection, and clean mapping could benefit from pre-segmenting the images before maneuvering or 3D reconstruction. We derive a training scheme that exploits depth maps of a reconstructed 3D model of a black smoker field in 1400 m water depth, resulting in a carefully selected, ground-truthed data set of 1000 images. Using this set, we compare the advantages and drawbacks of a classical Markov Random Field-based segmentation solution (graph cut) and a deep learning-based scheme (U-Net) to finding free space in forward-looking cameras in the deep ocean.

Actions (login required)

View Item