In context
Underwater robots operating near the seabed face two persistent challenges for vision-based navigation: sediment stirred up by thrusters obscures features, and artificial lighting creates uneven illumination. These conditions degrade the feature extraction and tracking that visual SLAM (simultaneous localization and mapping) relies on, limiting autonomous inspection and free-cruising missions. By early 2023, researchers were actively adapting terrestrial SLAM frameworks like ORB-SLAM to underwater conditions, but robust preprocessing remained an open problem.
What was reported
Researchers at Shanghai Jiao Tong University proposed an image preprocessing algorithm combining a semi-mean filter for dust removal and an illumination equalization method for feature enhancement. The semi-mean filter detects sediment particles by comparing window edge pixels to the global mean, then updates pixels from outside inward using adjacent already-filtered pixels, preserving background features while removing impurities.
For illumination, the algorithm assumes that under uniform, sufficient light, the same seabed terrain exhibits similar pixel distributions across different locations. Using this invariance, it estimates an attenuation coefficient and scattering weight to restore images to an evenly lit state, effectively enhancing features.
Tested on various underwater terrain datasets with ORB-SLAM3, the filtered and enhanced images increased the number of extracted feature points and mapping point clouds by an average of 200%, improving both performance and stability of the SLAM system.
Why it mattered
This work addressed a critical bottleneck in underwater visual SLAM, offering a practical preprocessing step that could make autonomous underwater vehicles more reliable for inspection, mapping, and exploration tasks. The approach's reliance on environmental similarity rather than deep learning made it adaptable to diverse underwater conditions without extensive training data.
The results show that the number of feature points extracted and the number of point clouds for mapping are increased by 200% in average by using the filtered and enhanced datasets.
Source: 《机器人》期刊 (robot.sia.cn) · Published 2023-03-11 · “水下视觉SLAM的图像滤波除尘与特征增强算法”