Visual place recognition for coastal scenes with semantic and sequential constraints

Visual place recognition for coastal scenes holds great significance for tasks such as automatic ship positioning, environment monitoring, patrols, and rescues. However, due to sparse features and complex environmental interference, existing visual place recognition methods designed for land scenes often yield suboptimal results when applied to marine environments. To address these challenges, this paper proposes a visual place recognition method that leverages semantic and sequential constraints. Firstly, we design a novel semantic constraint algorithm that guides the model to focus on landmark regions by distinguishing semantic categories within the scenes. Additionally, a horizon line correction technique is employed to mitigate the interference caused by changes in viewing angles during scene recognition. Moreover, a descriptor matching algorithm is introduced, which utilizes sequential constraints and hierarchical matching to better capture and adapt to temporal information. Through comprehensive training and testing using real-world coastal scene image sequences, the proposed method outperforms other state-of-the-art visual place recognition algorithms. It achieves a substantial improvement of 0.22 in first retrieve success rate compared to the second-best model, demonstrating its superior performance in coastal scene recognition.