【摘要】 现有的目标跟踪算法严重依赖于目标检测器的性能,如果目标检测器的虚警率或漏警率较高,数据关联将会失败,导致目标跟踪精度不足。为此,通过结构化学习为每个对象训练一个模型,将目标跟踪问题建模为拉格朗日松驰优化问题,提出一种目标身份感知网络流量(TINF)技术进行结构化学习的推理。在学习期间,通过搜索使目标身份感知网络流量代价函数最小化的一组轨迹,确定最被违反约束和序列在下个时间段的最优轨迹,推断视频片断中所有目标的最佳位置。利用多种高难度数据集进行仿真实验,实验结果表明,所提方法的性能优于其它较新算法。更多还原

【Abstract】 The existing target tracking algorithms rely heavily on the performance of the target detector,if the false alarm rate is higher,the data association will fail,which leads to the lack of the target tracking accuracy.For this reason,a model was trained for each object through structured learning,and the target tracking problem was modeled as a Lagrange relaxation optimization problem,and a target identity aware network flow technique was proposed for reasoning of structured learning.During the learning process,the target identity aware network traffic cost function was minimized by searching a set of trajectories,the most violated constraint and the optimal trajectory of sequence in the next time period were determined,and the best position of all the targets in the video clip was inferred.Experiments involved challenging yet distinct datasets,results show that the proposed method can achieve better results than the state-of-art ones.更多还原