基于CRF模型和标签代价的群组多目标跟踪算法

目标身份切换现象在目前的视频多目标跟踪算法中普遍存在,特别是在遮挡严重的场景中.针对这一问题,提出一种结合了CRF(condition random field)模型和标签代价函数的多目标跟踪算法.该算法将多目标跟踪问题转化为求解统一能量函数的最小解问题;同时,将目标的群组状态融合到跟踪器中,减少了目标发生身份切换的概率,提高了算法的鲁棒性.在多个公共数据集中对该算法进行仿真,实验结果显示,在多个性能指标特别是目标发生身份切换次数指标中,该算法优于目前主流的跟踪算法.

A group multi-object tracking method based on CRF model and label cost

ID switch is ubiquitous in present multi-object video tracking, especially it is almost inevitable in the heavy occlusion scene.To solve the problem a multi-object tracking method which combines the CRF model and label cost function is proposed and it transforms the multi-object tracking problem into an energy minimization problem. Meanwhile the group states are integrated into the tracker to decrease the probability of ID switch and improve the robustness of the method. The proposed method was used in some public datasets for simulation. Experimental results demonstrate that the proposed method outperforms the current classical methods on some performance indicators, especially the IDs indicator.