To address the shortcomings of existing simultaneous localization and mapping(SLAM)systems in semantic understanding and scene reconstruction

a neural implicit semantic SLAM system based on RGB-D sensors is proposed.This method uses a neural implicit representation approach to represent scenes

achieving multi-granular semantic understanding of scenes through hierarchical semantic representations.A cross-attention mechanism is employed to fuse appearance

geometry

and semantic features

improving the system’s robustness when single attributes are missing.An internal fusion decoder is designed to generate multimodal outputs of semantics

RGB

and truncated signed distance field(TSDF).Experiments on the Replica and ScanNet datasets show that compared to mainstream methods

this system achieves optimal tracking accuracy by 32%

reconstruction accuracy by 5%

and semantic segmentation by 10%

while also supporting real-time semantic mapping at 8Hz.