采用增量型非负矩阵分解建模的目标跟踪算法

An Algorithm for Object Tracking Based on Incremental Non-negative Matrix Factorization

建立鲁棒的外观模型是目标跟踪中的关键问题,为此提出一种基于增量型非负矩阵分解的目标跟踪算法.首先根据转移概率模型在当前帧中预测得到一组图像样本;随后利用非负矩阵分解获取样本在子空间中的坐标向量;在此基础上计算样本与前一帧视频中目标图像在低维坐标向量上的相关性,以具有最大相关性的图像样本作为目标在当前帧中的图像区域;最后以增量的方式完成子空间的在线更新,提高了外观模型的更新效率,且所要求的存储空间大小恒定.实验结果表明,该算法对目标物的外观变化具有良好的自适应性,能够在视频序列中对目标进行稳定的跟踪.

As it is crucial for object tracking to establish a robust appearance model,an algorithm for object tracking based on incremental non-negative matrix factorization is presented.Firstly,resorting to a transition probability model,a set of image patches are predicated as candidates for object image in the current frame,and then non-negative matrix factorization is used to obtain the low-dimensional coordinate vectors of the image patches.With the coordinate vectors,the associations between image patches and object image in the previous frame are evaluated,and the image sample with the maximum association is regarded as the image region of the moving object in the current frame.Finally,the subspace of object images is updated incrementally,thus the efficiency is improved in addition to a constant storage requirement.Experimental results show that our algorithm is able to adapt to variations in appearance of objects well,and it can track objects more steadily.