基于全局概率密度搜索的快速目标跟踪

 为了解决均值迁移目标跟踪算法中跟踪窗口对局部概率密度模式敏感的问题，该文提出一种基于全局概率密度搜索的目标跟踪策略。根据目标尺度设定一组从大到小排列的带宽序列，并依次根据每个带宽进行均值迁移迭代收敛，利用大带宽的平滑作用避开局部概率模式的干扰；依靠小带宽进行精确定位，最终收敛到真实目标区域。为了提高均值迁移的收敛速度，引入了Over-Relaxed优化策略加速迭代过程。在边界优化算法的收敛条件约束下，根据采用Over-Relaxed策略前后相关系数的变化，自适应地调整学习率。实验结果表明全局概率密度搜索能够有效地跟踪快速运动的目标，并且当目标短暂丢失时也有一定的恢复能力；Over-Relaxed策略也能显著的提高收敛步长，减少迭代次数。

Fast Object Tracking with Global Kernel Density Seeking

An object tracking algorithm with global kernel density seeking is proposed to avoid local probability mode in mean shift tracking process. Firstly, a monotonically decreasing sequence of bandwidths is obtained according to the object scale. At the first bandwidth, a maximum probability can be found with mean shift, and the next iteration loop started at the previous convergence location. Finally, the best density mode is obtained at the optimal bandwidth. In the convergence process, with the smoothness effect of the large bandwidth, the compact of the local probability mode is avoided, and the precise position of the object can be found with the optimal bandwidth, which is similar to the object scale. To speed up the convergence, Over-Relaxed strategy is introduced to enlarge the step size. Under the convergence rule, the correlation coefficient is used to adopt the learning rate. The experimental results prove that the proposed tracker with global kernel density seeking is robust in high-speed object tracking, and performs well in occlusions. The adaptive Over-Relaxed strategy is effective to lower the convergence iterations by enlarging the step size.