基于组合带宽均值迁移的快速目标跟踪

Fast object tracking with multi-bandwidth Mean Shift

为了解决传统均值迁移(Mean shift)目标跟踪算法中跟踪窗口容易收敛至局部概率模式的问题,提出一种基于组合带宽Mean Shift的目标跟踪策略,并建立了一种自适应学习率的over-relaxed优化策略以加速收敛过程。根据目标尺度设定了一组从大到小排列的带宽序列,并依次根据每个带宽进行Mean Shift迭代收敛运算,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,最终使其收敛到真实目标区域。由于组合带宽Mean Shift会造成一定的额外运算量,为此引入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-re-laxed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,组合带宽Mean Shift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;实验采用over-relaxed策略后,收敛次数减少了30%～70%。

An object tracking algorithm with multi-bandwidth and adaptive over-relaxed accelerated convergence was proposed to avoid the local probability mode in a Mean Shift tracking process. First- ly, a monotonically decreasing sequence of bandwidths was obtained according to the object scale. At the first bandwidth, a maximum probability could be found with the Mean Shift, and the next iteration loop started at the previous convergence location. Finally, the best density mode was obtained at the optimal bandwidth. In the convergence process, the compactness of the local probability mode was avoided with the smoothing effect of the large bandwidth, and the precise position of the Object could be found with the optimal bandwidth, which was similar to the object scale. To speed up the convergence, an over-relaxed strategy was introduced to enlarge the step size. Under the convergence rule, the correlation coefficient was used to adjust the learning rate adaptively. The experimental results prove that the proposed tracker with multi-bandwidth Mean Shift is robust in high-speed object tracking, and performs well in occlusions. The experimental results also show that the adaptive over-relaxed strategy reduces the convergence iterations by 30%-70 %.