I2KEN: Intra-Domain and Inter-Domain Knowledge Enhancement Network for Lifelong Loop Closure Detection

Loop closure detection (LCD) is crucial for robotic systems as it corrects accumulated pose estimation errors. Existing LCD methods mainly rely on data training of specific scenarios, which leads these methods to struggle to achieve the required cross-domain adaptability when deployed in unseen environments. Therefore, we propose Intra-domain and Inter-domain Knowledge Enhancement Network (I2KEN), a lifelong loop closure detection (LLCD) method, to enhance the cross-domain adaptability of LCD and prevent catastrophic forgetting. First, an Intra-domain Knowledge Refinement (IdKR) strategy is proposed, where a relation matrix among samples is constructed to correct feature relationships extracted by new and old models, effectively reinforcing accurate knowledge transfer. Second, we develop an Inter-domain Knowledge Consolidation (IdKC) module which consists of Batch-wise Contrastive Representation Distillation (BCRD) and Similarity-based Adaptive Model Fusion (SAMF), together enhancing anti-forgetting capability. Specifically, BCRD employs contrastive learning at the feature level to align new and old models, mitigating catastrophic forgetting. Finally, SAMF adaptively fuses the models based on their similarity relationships to balance performance across domains. Experimental results show that I2KEN exceeds existing LLCD methods in three large-scale datasets, achieving new state-of-the-art (SOTA) results for LLCD.