Abstract: In general, the multi-target, multi-frame data association needs to be exploited in multiple objects tracking. However, large number of samples and time complexity of the objective function limits its application. A novel real-time multi-object tracking algorithm is proposed. An on-line updating structure priors model is presented to capture the data space association. For the inference of space confidence, a dense sampling method is presented to overcome the imprecision of target state caused by traditional sparse sampling strategy. The surrounding regions of target are exacted as positive example, and then transformed circularly as negatives. Using the well-established theory of Circulant matrices, a fast learning method is achieved to infer the confidence for all possible position of target with the Fast Fourier Transform. Experimental results on a number of challenging sequences demonstrate the precision and efficiency of the method.

Key words: object tracking, structure priors, dense sampling, fast fourier transform

摘要： 多目标跟踪算法通常需要计算多帧、多目标间的数据关联，由于目标样本数量大，优化过程十分耗时，因此往往实际应用受限。提出一种实时的多目标跟踪算法，通过建立在线更新的结构先验模型约束目标间的空间位置关系，从而捕获多帧多目标间的数据相关性；在推理目标的空间置信度时，为克服传统方法使用稀疏采样造成样本不足引起目标状态估计不准确的问题，采用一种新的思路：提取目标及其周围区域作为正例样本，在计算过程中引入循环矩阵理论进行密集采样，并进一步通过对解进行傅里叶变换，实现对搜索窗口内所有样本似然的快速推理，从而为结构先验模型提供目标所有可能位置的置信度。实验结果表明了该算法在提高跟踪精度的同时显著降低了运算时间。

关键词: 目标跟踪, 结构先验, 密集采样, 傅里叶变换