Systems analysis of feature extraction and matching methods in opt acoustic data for navigation and mapping tasks

The scattering of light in turbid water and the presence of multipath acoustic reflections generate fundamentally different distortions in optical and acoustic images, which significantly complicates their direct interaction in un-derwater navigation and mapping tasks. A multimodal approach is proposed that combines visual and acoustic data in a single architecture of localization and map construction. The developed algorithm includes the formation of a common latent space of scene features based on contrastive self-learning applied to data from a stereo camera and an overview multipath sonar. The obtained descriptors are supplemented with odometric information from the inertial measurement unit and speed measurements from the Doppler lag, which allows us to construct a factor graph and provide a consistent assessment of the trajectory and dense map of the environment. The key component is a modified neural network architecture for detecting key points, trained on synthetic data from a digital twin in a simulator environment, followed by further training on real underwater scenes. To increase the generalizing ability, the homographic adaptation method was used, taking into account epipolar constraints and the probabilistic model of sensor noise. Experiments conducted in a laboratory pool and in an open water environment demonstrated an increase in the accuracy of feature matching, a decrease in the average squared localization error, as well as a decrease in the loss of decorrelation of key points. The results obtained confirm the applicability of the proposed method in conditions of limited visibility to increase the reliability of simultaneous localization systems and the construction of three-dimensional maps of the environment.