Researchers from the University of Michigan have developed a novel framework to enhance underwater robot perception, addressing key challenges in stereo depth estimation and Simultaneous Localization and Mapping (SLAM). The team, led by Ram Vasudevan and Katherine A. Skinner, has introduced SurfSLAM, a system designed to improve real-time SLAM in underwater environments by integrating stereo cameras with IMU, barometric, and Doppler Velocity Log (DVL) measurements.
Underwater robots face significant hurdles in localization and mapping due to the unique challenges posed by submerged environments. Stereo cameras, while offering a cost-effective solution for depth estimation, struggle with image degradation caused by light attenuation, visual artifacts, and dynamic lighting conditions. These issues are exacerbated by the lack of rich texture in underwater scenes, making it difficult for stereo estimation networks trained on above-water data to perform accurately underwater. Additionally, the scarcity of real-world underwater stereo datasets hampers the supervised training of neural networks.
To overcome these obstacles, the researchers proposed a novel framework that enables sim-to-real training of underwater stereo disparity estimation networks. This approach involves using simulated data for initial training, followed by self-supervised fine-tuning with real-world data. By leveraging learned depth predictions, the team developed SurfSLAM, which fuses stereo camera inputs with additional sensor data to achieve real-time SLAM. This integration enhances the accuracy of trajectory estimation and 3D reconstruction in complex underwater environments, such as shipwreck sites.
The researchers collected a comprehensive dataset featuring over 24,000 stereo image pairs, accompanied by high-quality photogrammetry models and reference trajectories. This dataset was used to evaluate the performance of SurfSLAM and demonstrate its advantages in improving stereo estimation and enabling precise 3D reconstruction. Through extensive experiments, the team showcased the effectiveness of their training approach in real-world underwater scenarios, highlighting its potential to advance underwater robot perception and navigation.
The development of SurfSLAM represents a significant step forward in the field of underwater robotics. By addressing the fundamental challenges of stereo depth estimation and SLAM, this research paves the way for more accurate and reliable underwater exploration and mapping. The integration of multiple sensor inputs and the use of simulated data for training offer a robust solution that can be applied to various underwater robotic applications, from scientific research to marine archaeology and environmental monitoring. Read the original research paper here.