<p>Radar detection systems often face challenges with ghost targets, especially when targets are in motion. Existing methods underutilize the multidimensional information in radar point clouds, relying on point-wise classification that lacks overall semantic labels of the targets, leading to frequent misclassifications in complex environments. Moreover, current methods perform poorly in tracking sparse radar point clouds, especially in scenarios with crossing motions of targets. To address these issues, we propose SemanticTrack, a moving target semantic detection and tracking method. During the target detection stage, the proposed Moving Target Extraction module and Overall Semantic Segmentation network effectively exploit the multidimensional information, enhance the overall semantic comprehension of targets, and avoid the misjudgment problem. Additionally, the proposed Track Management module in the tracking stage combines temporal information and semantic labels to enable precise identification and suppression of ghost targets. Experimental evaluations on a self-collected dataset demonstrate that SemanticTrack achieves the highest ghost suppression rate of 93.63%, significantly outperforming existing methods. In terms of tracking performance, compared to the classical AB3DMOT, it achieves a 32.26 percent points improvement in multi-object tracking accuracy, while reducing multi-object tracking precision by 0.921 m. In particular, this method exhibits superior stability in multi-target cross-motion scenes.</p>