Adaboost检测和混合粒子滤波融合的多目标跟踪

Multi-target tracking of Adaboost detection combining with hybrid particle filtering

针对多目标跟踪时因存在很多不确定性因素,而导致粒子滤波不能有效处理多模式的增长问题,首先,提出一种混合粒子滤波跟踪方法,通过混合权值的计算实现粒子间的相互关联,能有效地保持和处理多模式问题;其次,为了提高算法对数目变化的多目标跟踪处理能力,在混合粒子滤波跟踪算法中,又融入了Ada-boost检测算法,用动态模型和Adaboost检测信息合并成的混合观测模型构造似然函数,实现了一种能学习、检测和跟踪感兴趣目标的跟踪系统;最后,在刚性、非刚性以及数目变化的多目标视频序列中对算法进行测试,结果表明算法对数目变化的多目标能实现有效跟踪.

To reduce the uncertainties of multi-target tracking,a hybrid particle filter method was proposed.The interconnection was realized among particles through the calculation of hybrid weights,which could effectively maintain and deal with multi-mode problems.Secondly,in order to improve the mul-targets tracking processing ability of the algorithm when the number of multi-target changed,the Adaboost detection algorithm was integrated into hybrid filter tracking algorithm.Likelihood function was constructed through hybrid measurement model which was merged between the dynamic model and Adaboost detection.Then a tracking system to learn,detect and track interesting target was realized.Finally,the algorithm was tested in multi-target video sequence of a rigid,non-rigid and number-changing.The experiment shows that the algorithm can effectively track multi-targets changing in numbers.