GGIMN: Gravity Gradient Image Matching Network With Adaptive Adjoint Smoothing for AUV Navigation

Gravity gradient-aided navigation is a promising navigation method for autonomous underwater vehicle (AUV), in which matching algorithm is its key technology. Image matching algorithms have higher matching accuracy and information utilization than single point matching and sequence matching. However, the spatial domain image feature matching ability is weak in areas with insignificant feature changes due to the uneven feature distribution in the gravity gradient reference map. To break through this limitation and solve the problem of insufficient spatial domain image feature matching ability, we propose a gravity gradient image matching network (GGIMN) with adaptive adjoint smoothing for AUV navigation and positioning. First, a training set for GGIMN with equal numbers of positive and negative samples is constructed by preparing the gravity gradient reference map and fitting the gravity gradient measured map. Then, the feature extraction part of MatchNet is expanded to construct a deep learning network with five gravity gradient twin towers in parallel. Subsequently, an adaptive adjoint smoothing method that combines the idea of sequence matching is proposed to solve the problem of unsmooth matching trajectory and further improve the matching accuracy. Finally, the GGIMN is trained offline, and the gravity gradient map is matched and positioned online. The results of marine experiments show that the GGIMN has higher matching and positioning accuracy, and the average positioning error is reduced beyond 70% compared with one-dimensional matching algorithms such as single-point iteration and sequence matching.