In complex dynamic scenarios characterized by crowding and occlusion, existing lightweight multi-object trackers often suffer from identity switches and trajectory fragmentation due to insufficient feature discriminability. To break the trade-off between high accuracy and low computational cost, this study builds upon the advanced end-to-end tracker MOTIP and proposes a novel collaborative multi-object tracking algorithm, distillation-driven and query-separated tracker (DQS-Tracker). This algorithm improves lightweight models from two complementary perspectives: knowledge and architecture. First, an adaptive hierarchical knowledge distillation scheme tailored for multi-object tracking is proposed. By aligning visual features, identity-discriminative logits, and temporal consistency representations, the scheme transfers the expert decision-making capabilities of the large teacher model in complex scenarios to the student model, while an adaptive weighting strategy is introduced to optimize training dynamics. Second, a query duty separation mechanism is proposed to partition detection queries into track and newborn queries, thus reducing attention competition between tracking and detection tasks. Experimental results on two challenging benchmarks, DanceTrack and SportsMOT, demonstrate that DQS-Tracker achieves performance comparable to that of the large teacher model on key metrics such as HOTA and IDF1, validating the generalization ability and robustness of this method across diverse complex scenarios.