基于多空间显著特征的目标跟踪算法

Object Tracking Algorithm Based on Discriminative Feature in Multiple Sub-spaces

在复杂场景中,目标的表示常常受到视角变化、光照变化、遮挡变化和背景干扰等影响。针对复杂环境下单一空间中表观模型的不鲁棒造成的跟踪不准确问题,在粒子滤波跟踪框架下提出一种基于多空间显著特征的目标跟踪算法。首先,针对单空间中目标表示的不准确性,通过各个子空间中目标和背景的差异性分析,进行显著空间的筛选。其次,在显著空间筛选的基础上,基于特征的差异性,构建目标的显著子空间矩阵表示,提升目标表示的鲁棒性。最后,通过粒子的显著表示和空间结构进行样本的有效筛选,降低冗余样本的干扰,提升跟踪的准确性。实验结果表明,在具有挑战性的视频上与现有流行的跟踪算法进行比较,证明了显著特征的有效性,且算法具有较好的鲁棒性和准确性。

In complex scene,the representation of objects is often affected by changes in perspective,illumination,occlusion and background interference.Aiming at the inaccuracy of tracking caused by the non-robustness of the apparent model in a single space in complex environment,an object tracking algorithm based on discriminative feature in multiple sub-spaces is proposed in the particle filter tracking framework.Firstly,for the inaccuracy of the target representation in a single space,the discriminative space is filtered by the difference analysis of the target and background in each sub-space.Secondly,on the basis of discriminative sub-space selected and feature difference,the discriminative sub-space matrix representation of the target is constructed to improve the robustness of the target representation.Finally,through discriminative representation of particles and spatial structure,effective screening of samples can be carried out to reduce the interference of redundant samples and improve the accuracy of tracking.The experiment shows that compared with the existing popular tracking algorithms in challenging video,the effectiveness of the discriminative features is proved,and the algorithm has better robustness and accuracy.