基于EM的多目标跟踪算法

为了提高多目标跟踪的鲁棒性,增强目标之间的区别性,使用了一种基于能量最小化(energy minimization,EM)的多目标跟踪算法,不同于现有算法,本算法专注于将多目标跟踪中的复杂问题表示为能量函数的模型,模型中包括了更优的目标区分策略(相似度模型)。通过将每个能量函数成本值对应一个多目标的跟踪轨迹方案,算法将多目标跟踪问题转化为能量最小化的问题。在能量函数模型的优化方法上,算法采用共轭梯度算法和一系列的跳转运动来找到能量最小的值。公开数据集的实验结果证明了本算法的有效性,而且定量分析结果证明了本算法提高了目标与背景、目标之间的相互区别性从而与其他算法相比能获得更好的鲁棒性能。

Multi target tracking algorithm based on EM method

In order to improve the multi-target tracking robustness and enhance the difference between the targets,this paper uses an energy minimization method for multi-target tracking.Different to the existing algorithm,the algorithm focuses on the representation of the complex problem in multi-target tracking as energy function model,which includes a better target segmentation strategy (similarity model).By assigns every possible solutions a cost (the "energy"),the algorithm transforms the multiple target tracking problem into an energy minimization problem.In the energy minimization optimization method,the algorithm uses the conjugate gradient algorithm and a series of jump moves to find the minimum energy value.The experimental results of open data demonstrate the effectiveness.And the quantitative analysis results show that this algorithm can improve the difference between targets or between target and background so as to obtain better robust performance compared with other algorithms.