基于多阶段关联的多目标跟踪算法

针对现有多目标跟踪算法关联过程中，外观和几何信息利用不充分，同时跟踪对象的邻域间信息交互不足的问题，提出了一种基于多阶段关联的多目标跟踪算法，根据目标之间的不同关联状态，将几何信息和外观信息合理应用于不同关联阶段。算法提出了基于正则化距离交并比(DIoU-Mea)的匹配模块，仅利用几何信息快速将强关联目标匹配。同时基于稀疏图网络(GNN)的关联模块对跟踪对象的邻域建模，促进对象之间的信息交换并提高跟踪精度。基于通道注意力融合特征模型和形状交并比的双校验模块(Double-Revise)进一步细化跟踪结果。所提算法利用不同阶段匹配算法的互补优势，在各阶段合理利用外观和几何信息，过滤掉错误的匹配并识别正确的目标对应关系，在MOT17数据集上进行了验证与测试，其高阶跟踪精度(HOTA)在测试集中达到了64.8%,表明算法具有较好的性能，在密集场景下具有较好的鲁棒性。

Multi-object tracking algorithm based on multi-stage association

Existing multi-object tracking algorithms make insufficient use of appearance and geometric information, and the informationexchange among adjacent regions of the tracked object is limited. To solve this problem, a multi-object tracking algorithm based onmulti-stage association is proposed, which applies geometric and appearance information to different association stages according to thedifferent association states among objects. Firstly, a fast matching module based on the regularized distance intersection and union ratio(DIoU-Mea) is employed to efficiently handle the matching task of strongly correlated objects only using geometric information.Secondly, an association module based on the sparse graph network (GNN) is incorporated to model the neighborhood of the trackedobject, facilitate information exchange among objects, and improve tracking accuracy. Finally, a double verify module (Double-Revise)is introduced, which utilizes the channel attention fusion feature model and the shape intersection and union ratio to further refine thetracking results. By utilizing the complementary advantages of different stage matching algorithms and making reasonable use ofappearance and geometric information in each stage, the proposed algorithm effectively filters out incorrect matches and accuratelyidentifies the correct object correspondence. The proposed algorithm is evaluated and tested on the MOT17 dataset. Its high-ordertracking accuracy (HOTA) reaches 64. 8% on the test set. Results show its good performance and robustness in dense scenarios.