稀疏子空间的粒子滤波法在目标跟踪中的应用

Application of sparse subspace based particle filtering in target tracking

针对目标跟踪算法中对光流变化较敏感、对目标的内蕴变形以及外观变化鲁棒性低等问题,将压缩感知理论及稀疏子空间表示方法应用于目标跟踪。对跟踪目标压缩采样,根据压缩感知原理采样后的信号在较高的概率下保持原有的信息量的特点,为目标跟踪提供足够的信息保证,利用鲁棒主成分分析(robust PCA,RPCA)方法提取压缩采样目标的稀疏主成分;在粒子滤波理论框架下实现跟踪过程,利用过去连续时间序列下目标的状态估计当前时刻目标的状态;在对目标所在的稀疏子空间更新的处理中,利用子空间与候选目标的相似性更新子空间。实验结果表明,该算法具有较强的鲁棒性,较好地满足了实时性要求。

To improve target tracking robustness to challenges such as sensitivity to optical flow,changes of appearance which used to model the target,an efficient tracking algorithm composed of compressed sparse representation of the target and particle filtering was proposed.The object was sampled compressively with the theoretical guarantee that the original signals can be preserved with high probability,which was enough for target tracking.The sparse subspace was extracted from the compressed images with RPCA.The particle filtering was used to achieve tracking process,namely,the target state was predicted with the bypast information in new frame.Sparse subspace was updated through measuring the similarity to the candidate target.Experimental results show that the proposed algorithm is robust and meets real-time request.