| Underwater terrain matching localization technology uses underwater terrain features as reference information to determine the position of underwater vehicles.Due to its characteristic of positioning error not accumulating over time,it can provide errorbounded position correction for underwater vehicles during long-term diving missions.The terrain matching algorithm is a key factor that affects the performance of underwater terrain matching and localization.Since underwater terrain information can be represented in image form,image processing techniques can be used to obtain the position information of underwater vehicles.However,factors such as the relatively slow changes in underwater terrain,low precision of underwater terrain measurement sensors,and the variability of vehicle attitude,lead to weak textures,nonlinear intensity differences,and relative rotations between the underwater terrain images obtained before and after,posing great challenges to the matching algorithm.Based on the current research progress in image matching,this paper conducts relevant research on the key factors affecting terrain image matching,and aims to construct a high-precision and robust underwater terrain image matching algorithm to improve the stability and reliability of underwater terrain matching and localization.The main research contents and related works of this paper are as follows:Firstly,the method of image matching for coarse-grained reference terrain is studied.In response to the problems of smooth grayscale variations and insufficient representation of terrain details in coarse-grained reference terrain images,a method based on salient superpixels for underwater terrain image matching is proposed.Since the density of contour lines can reflect the amount of terrain variation,this method utilizes the feature aggregation capability of superpixels and their responsiveness to edges to extract contour lines from the terrain images.By quantifying the richness of the terrain within the neighborhood of superpixels,dense regions of contour lines are selected for description and matching,thereby enhancing the expressive power and stability of the feature vectors.Furthermore,to address the issue of rotation sensitivity in the feature vectors,a method is proposed to ensure the rotation invariance of features by computing the main direction of the superpixel neighborhoods.Experimental results demonstrate that this method can adapt to the matching tasks of coarse-grained terrain images and exhibit robustness and effectiveness in dealing with nonlinear intensity differences and planar rotation variations between images.Secondly,the method of image matching for small undulating terrains is studied.In response to the weak texture characteristics and sensitivity to nonlinear intensity differences in small undulating underwater terrain images,a hierarchical matching strategybased underwater terrain matching method is proposed using a supervised algorithm.To address the issue of insufficient training data for underwater terrain images,a data generation strategy based on candidate region sample selection and an environment simulationbased data augmentation strategy are proposed to construct an underwater terrain dataset by simulating batch matching methods and environmental factors.To deal with the nonlinear intensity differences and rotation variations in underwater terrain images,a hierarchical matching strategy that combines deep and shallow features is proposed based on a dual-branch network,which improves the anti-interference capability against false matching regions.Experimental results demonstrate significant performance improvement of this method compared to traditional image matching methods based on handcrafted features in underwater terrain matching tasks.Thirdly,the method of image matching for highly self-similar terrains is studied.In response to the difficulty in distinguishing regions with high local self-similarity in small undulating terrains,a contrastive learning-based underwater terrain image matching method is proposed.This method alleviates the influence of nonlinear intensity differences through data augmentation.To address the issues of limited samples and challenging annotations,a position-based positive-negative sample generation approach is proposed,and the problem of center position bias in the network is resolved.Furthermore,a self-attention mechanism is introduced to fuse contextual features,enhancing the expressiveness of the features.The matching problem of multi-scale templates is addressed by introducing a masking mechanism.The model is trained in a self-supervised manner through data contrast,improving the discriminative power of the features.Experimental results demonstrate that the generated feature representations have higher discriminability and can achieve precise matching in areas with relatively flat terrain or high self-similarity.Finally,the method of image matching for large rotation scenes is studied.To address the sensitivity of batch terrain matching algorithms to angle rotation,a rotation-invariant underwater terrain image matching method based on circular mask attention network is proposed.Firstly,considering the inability of typical feature networks to encode image rotation changes,a rotation-equivariant feature network is constructed to extract rotationequivariant features from the images.Then,considering the non-differentiability of circular projection transformation,a modularization approach is proposed using circular masks and attention mechanisms to modularize the circular projection transformation,which is embedded as a layer in an end-to-end network model together with the rotation-equivariant feature network,further encoding the rotation-equivariant features into rotation-invariant features.The method combines the feature expression capability of learning-based methods with the rotation invariance of circular projection transformation,leveraging knowledge from both domains.Experimental results demonstrate that the proposed method improves the robustness and matching accuracy of underwater terrain images in the presence of relative rotations. |
