Purpose This study aims to address the degradation of localisation and mapping accuracy in visual simultaneous location and mapping (SLAM) caused by moving objects in dynamic environments. This study proposes a lightweight SLAM method based on a hierarchical dynamic-object discrimination mechanism. Design/methodology/approach A three-layer pipeline is established: a pruned and quantised YOLOv8n-seg performs real-time instance segmentation to remove features of inherently dynamic categories; dense scene flow combined with rigid-motion consistency and statistical criteria identifies potentially dynamic objects among semantically static ones; during back-end optimisation, dynamic observation factors are decoupled, while a progressive feature management strategy maximises the retention of static constraints. Findings Evaluations on highly dynamic TUM RGB-D sequences (fr3\walking\ₓyz, fr3\walking\ₕalf) demonstrate that the proposed method outperforms ORB-SLAM3, DynaSLAM and DS-SLAM, achieving ATE/RPE of 0. 014 / 0. 017 m and 0. 019/0. 023 m, respectively, confirming its robustness, accuracy and real-time performance. Originality/value This study proposes a novel lightweight hierarchical framework that integrates efficient semantic segmentation, motion-aware geometric discrimination and dynamic-factor decoupling in optimisation. It offers a practical and accurate solution for robust visual SLAM in real-world dynamic environments.
Cao et al. (Mon,) studied this question.