Abstract: Aiming at the problems of significant cumulative SLAM positioning errors and prolonged relocalization time caused by the sparse and highly repetitive environmental features in coal mine underground environments, this paper proposes a multi-sensor mapping and localization technology integrated with semantic landmarks. By establishing a mapping relationship between roadway structural features and landmark encoding information, an underground semantic map incorporating spatial geometric features and custom semantic labels is constructed. Utilizing landmarks for initial pose calculation effectively suppresses interference from repetitive roadway features and substantially reduces relocalization time. Meanwhile, dynamic correction of robot poses achieved through semantic landmarks significantly eliminates cumulative SLAM errors. To validate system performance, experimental scenarios were set up and deployed on a dust-flushing robot. The results demonstrate an average mapping error of 2.0 cm, static positioning errors below 3.5 cm, dynamic positioning errors below 15.3 cm, and an average relocalization time of 3.3 seconds.