Multi-Object Tracking (MOT) has been one of the most important topics in computer vision. The tradi-tional tracking-by-detection framework of MOT is severely suffered from the poor detection results. In this paper, based on Transformer, we introduce the tracking-by-query MOT framework, and propose to apply semantic segmentation as an auxiliary task to optimize the training of MOT trackers, which addresses more on extracted foreground features. In addition, a feature-dependent dynamic object query (DOQ), instead of a fixed-learned object query (LOQ), is put forward to retrieve the new detections, improving the flexibility and constringency of the framework. We tested our SegDQ method on various scenarios including MOTChallenge 15, 16 and 17 datasets. The experimental results show that it obvi-ously improves the MOTA and IDF1 indexes of tracking results. (c) 2022 Published by Elsevier B.V.