Multi-Object-Tracking (MOT) methods have made great efforts on discovering more differences among objects. However, the correlation between objects is rarely discussed which is also important. In this paper, we propose two modules considering their correlation in graph to enhance association accuracy and help motion prediction for MOT methods with Tracking-By-Detection (TBD) paradigm, where the relative positions between objects are represented as graph edges. By ensuring the structure of graph in continuous tracking and regarding the objects in occlusion as a whole, these two modules correct the wrong assignments and predict motions for occluded objects aided by their correlated objects. After inserting our modules into an existing TBD method, the improvements in benchmark results and ablation study of three MOT datasets demonstrate their effectiveness.