粒子群算法是一种新的进化算法, 算法思路适合于进行视频跟踪, 但是由于在视频跟踪过程中以跟踪窗口作为粒子, 因此该粒子具有中心点横坐标、中心点纵坐标和窗口半径三维特征向量, 计算冗余较大, 难以满足视频跟踪的实时性要求。提出了一种多粒子群视频跟踪算法, 即在跟踪过程中使用多个粒子群, 粒子群与粒子群之间粒子半径不同, 在各粒子群以评价函数收敛到最佳中心点后, 再完成各自半径的一维粒子群计算。这样就可将三维粒子群计算分为一个两维和一个一维粒子群计算, 最后通过比较得出最佳粒子, 作为搜索结果。分析了这一算法成立的必要条件, 即当选择Bhattacharyya 系数计算方法作为粒子群算法的评价函数时, 大于目标的固定窗体的中心点可以收敛到目标的形心。实验证明, 这种基于多粒子群的跟踪算法可以应用于实时视频跟踪, 其跟踪效果优于传统算法。

The particle swarms algorithm is a new evolutionary algorithm, its ideal is suitable for video tracking, but since the tracing window is considered as particle in the process of video tracking, so this particle has three-dimensional characteristic vectors i.e. the vertical and lateral coordinate of the central point, the radius of window, the calculated quantity is too large to meet the real time requirement in video tracking. The changes of object scale and position whithin the kernel will not impact localization accuracy of particle swarms based on tracking algorithm using the Bhattacharyya coefficient as criterion function. According to this analysis, a multiple particle swarms algorithm of video tracking is presented, the ideal is that several particle swarms are used in tracking process, the particle radius between particle swarms is different, but the particle radius in the same swarm is the same, so that there is only two-dimensional motion while every particle swarm are searching its optimum central point, the factor of variation (mutation) is introduced in the process in order to let one particle swarm can be under the influence of the other particle swarms. After each particle swarm convergences to its optimum central point according to criterion function with Bhattacharyya coefficient, the one-dimensional particle swarm calculation for each radium can be completed, it can be seen that the three-dimensional particle swarms calculation is divided into one two-dimensional and one one-dimensional warms calculations in the new method. At last the results of particle swarms calculation are compared and the optimum particle is obtained as the searching result. The simulation results show that the presented multiple particle swarms algorithm of video tracking is able to apply in real time, and the tracking performance is superior to conventional algorithms.