SMR-GA: Semantic Map Registration Under Large Perspective Differences Through Genetic Algorithm

摘要

The registration of multiple local maps is a crucial step in multi-robot collaboration. However, existing methods are insufficient to address the complexities associated with sparse features and elevated outlier rates, primarily due to significant discrepancies in perspective. This paper proposes a semantic map registration method using a genetic algorithm to handle larger perspective differences. We introduce a potential corresponding point selection strategy to eliminate outliers. A fitness calculation model has been developed that closely integrates semantic and geometric features. To assist in avoiding local minima and accelerating convergence, an adaptive mutation step adjustment strategy is proposed. Experiments on multiple datasets demonstrate that our algorithm significantly improves accuracy and success rates. Successful registration can still be achieved even when the perspective difference exceeds 90°. Ablation experiments confirm the effectiveness of each model.