Dynamic objects in LiDAR SLAM often introduce ghosting artifacts that degrade map quality. While offline methods can successfully clean these maps, they lack real-time capabilities. Conversely, online methods often suffer from state oscillation (where moving objects are misclassified as static when they temporarily stop) and incomplete point cloud removal. To address these challenges, we propose DGMSE, a real-time framework for removing dynamic point clouds in complex urban environments. Our approach consists of three sequential steps. First, the PointPillars 3D detection network quickly isolates potential dynamic objects, significantly reducing computational overhead. Second, to mitigate state oscillation, we introduce a five-state Markov chain to continuously estimate the probabilistic motion state of objects, which is not limited to simple moving-or-static classifications. This temporal memory prevents misclassification during sudden stops or motion transitions. Finally, instead of simply deleting all points inside a detection box, a density-adaptive clustering approach evaluates the geometric and motion characteristics of points, precisely segmenting dynamic objects from the static background without damaging permanent structures. Evaluated on the SemanticKITTI dataset, DGMSE achieves an F1-score of 0.95–0.99, achieving performance competitive with offline methods while satisfying real-time processing requirements.