Dense Semantic SLAM Using TSDF Mapping and Deep Learning-Based Loop Closure

This paper proposes a robust Semantic SLAM system designed to generate high-fidelity maps for indoor environments. Addressing the drift problem in long-term navigation, we introduce a deep learning-based loop closure module that utilizes GeM (Generalized-Mean) pooling for global place retrieval and LoFTR (Detector-Free Local Feature Matching) for precise geometric verification. This approach allows for globally consistent trajectory optimization, effectively mitigating cumulative odometry errors. Based on the optimized poses, depth measurements are fused into a TSDF volume to reconstruct a dense surface enriched with semantic information projected from 2D segmentation masks. Experimental results demonstrate that our approach effectively minimizes drift, resulting in more robust global consistency compared to standard frame-to-model tracking.