HCSG-SLAM: Lightweight History-Consistent Semantic Fusion for Gaussian SLAM with Noisy Semantic Predictions

Online semantic Gaussian SLAM typically updates three-dimensional semantic maps using frame-wise predictions from two-dimensional semantic segmentation networks. However, occlusions, viewpoint changes, and object boundaries may introduce transient misclassifications, which can further cause cross-frame semantic inconsistency and error propagation. To address this issue, we propose HCSG-SLAM (History-Consistent Semantic Gaussian SLAM), a lightweight history-consistent semantic fusion method for Gaussian SLAM that introduces a compact Gaussian-level historical semantic state without modifying the main RGB/depth tracking and reconstruction pipeline of SGS-SLAM. Each Gaussian maintains only a candidate semantic label, a voting state, and an effective historical observation count. A streaming voting strategy forms the historical semantic consensus, while a strictly causal target-before-current-update strategy prevents the current observation from contributing to its own correction target. Experiments on Replica and ScanNet used segmentation-network predictions rather than ground-truth semantics as input. On Replica, the mean mIoU increased from 0.300830 to 0.307010, with positive gains in all 15 paired runs. On ScanNet, the evaluation was extended to five scenes, where the overall mean mIoU increased from 0.366279 to 0.372166, with positive gains in 22 out of 25 paired scene–seed runs. Per-class analysis shows that the gains were concentrated in several structural and furniture categories, while a small number of classes remained unchanged or slightly decreased. Positive gains were also observed under all nine controlled transient-corruption settings, and additional persistent class-confusion and structured boundary-shift tests yielded positive or comparable performance across the evaluated scenes. Compared with FD-EMA, HCSG-SLAM achieved comparable semantic accuracy while requiring only 12 B of additional temporal state per Gaussian compared with approximately 202 B for FD-EMA. HCSG-SLAM introduced approximately 2.05% additional mapping time relative to the Baseline while leaving tracking essentially unchanged. These results indicate that compact Gaussian-level temporal semantic states can improve semantic consistency and robustness with modest online fusion overhead.