BPC-SLAM: Part-Level Dynamic Suppression and Structure-Constrained RGB-D SLAM for Human-Centric Dynamic Environments

To address the tracking degradation caused by global semantic-mask removal in human-dynamic scenes—where stable observations are mistakenly discarded, leading to insufficient geometric constraints—we propose a part-level motion determination strategy that suppresses truly dynamic interference while preserving valid observations from relatively static human parts, and further improves the robustness and accuracy of pose estimation by incorporating human structural constraints. Accordingly, we present BPC-SLAM, an RGB-D SLAM system featuring part-level dynamic suppression and structure-constrained optimization for dynamic human environments. Built upon ORB-SLAM3, the proposed system introduces human instance segmentation and pose keypoint estimation in the front end to partition each human instance into semantic parts such as the head, torso, and limbs. We then perform part-level robust statistics on the residual of dense optical flow after camera-motion compensation, so that only genuinely moving parts are suppressed while geometrically informative observations from relatively static parts are retained, thereby alleviating over-rejection in low-dynamic or locally moving scenarios. Furthermore, in the back end, constraints including the human–camera relative pose, keypoint reprojection, temporal motion priors, and bone-length consistency are integrated into a unified graph optimization framework to enhance observability and robustness under dynamics. Experiments on the TUM RGB-D dataset show that, compared with ORB-SLAM3, our method reduces ATE-RMSE by 97.38% on the highly dynamic sequence fr3/walking and by 46.79% on the low-dynamic sequence fr3/sitting. Additional comparisons, ablation studies, and real indoor-scene tests further demonstrate the effectiveness and stability of the proposed approach.