In unmanned driving scenes， dynamic objects can significantly affect the global accuracy and robustness of simultaneous localization and mapping （SLAM） systems. Most existing laser SLAM systems are prone to odometry drift， positioning failure， and mapping ghosting in dynamic environments. To address these issues， this paper proposed a semantic laser SLAM system for dynamic scenes that integrates a lightweight PointPillars target detection network and a multi-object tracking method. The system first uses the PointPillars network to obtain the bounding boxes of potential dynamic targets and filters the feature points within these bounding boxes to obtain a preliminary estimation of odometry. Secondly， the system constructs a temporal sliding window based on the tracking results obtained by the multitarget tracking algorithm， which is based on the uniform velocity Kalman Filter. This enables the establishment of a robust and efficient spatiotemporal association of target-level data， which removes dynamic objects and recovers static targets， further optimizing the odometry. Finally， the proposed method is compared with the state-of-the-art methods on KITTI and NUSCENES datasets in challenging dynamic environments. The experimental results demonstrate that our system significantly improves the accuracy and robustness of odometry and global mapping， while the system also maintains real-time performance，which meets the requirements of autonomous robot systems and intelligent transportation applications in dynamic scenes.