Abstract: Translating multi-target tracking into graphs is a problem as a whole. First of all, the application of classic layered thinking is adopted, and two tracking frameworks are established, and then target movement characteristics and appearance characteristics blend into the weight, which can simulate real tracking scene more accurately. Then, the dummy node is joined to deal with missing targets problem, which gives its accelerated version： aggregated dummy nodes. Finally, the mixed-binary-integer programming is used to solve the undirected graph to obtain a series of clusters at the same time. Experimental results on common dataset show that the proposed algorithm can realize real-time tracking and the tracking result is better.

Key words: multi-target tracking, graphs, aggregated dummy nodes, mixed-binary-integer programming

摘要： 整体上将多目标跟踪问题转化为图的问题。首先采用经典的分层思想，建立两层跟踪框架，并将目标的运动特征和外观特征融入权值，可以较精确地模拟真实的跟踪场景。接着，加入虚拟结点以处理目标缺失的问题，并给出其加速版：聚合虚拟结点。最后利用最大二值整数规划求解无向图以同时获得一系列团。实验在公共数据集上进行，结果表明，该算法可以实现实时跟踪，且跟踪结果较好。

关键词: 多目标跟踪, 图, 聚合虚拟结点, 最大二值整数规划