Multi-scale Semantic Feature Detection based Loop Closure Detection for Visual SLAM

In long-term, large-scale navigation and localization, reducing cumulative errors from robot motion is vital for global map optimization. Traditional loop closure detection methods, often based on the bag-of-words (BoW) model, face limitations such as manual feature design, long training time, operator dependence, and poor robustness under varying lighting. This paper proposes a loop closure detection method using multi-scale semantic information. High dimensional semantic features extracted by StarNet are first reduced via UMAP, preserving key information while removing redundancy. This nonlinear reduction captures complex relationships in dynamic scenes. PatchNetVLAD is then applied to enhance spatial feature representation by aggregating local features, improving similarity detection. Finally, weighted similarity scores are fused with the BoW model to achieve high-precision loop closure detection. Experiments on the New College and City datasets show the proposed method outperforms state-of-the-art approaches in various scenarios.