针对子空间表示跟踪算法处理遮挡问题的能力不足, 以及稀疏表示跟踪算法无法满足跟踪实时性要求等问题, 本文提出一种稀疏正则约束的子空间视觉跟踪算法。 该算法结合了子空间表示与稀疏表示的优势, 提升了对于遮挡问题的处理能力, 并且降低了算法的计算复杂度。首先, 该算法利用PCA子空间基向量集、子空间均值以及表示残差对目标进行表示, 同时算法采用L2范数作为表示系数以及表示残差的稀疏约束函数。其次, 算法采用了一种分步循环迭代的方法求解表示模型的系数与残差。然后, 为了保证子空间基向量与空间均值能够持续准确的描述目标在跟踪过程中出现的变化, 算法根据经过开运算处理后的表示残差中非零元素的不同比率构建不同的更新模板, 并结合增量主成分分析方法在线学习新的基向量与均值。最后, 在实验部分, 本文将提出算法在10个实验序列上的跟踪结果与8个现今主流跟踪算法进行对比, 同时从定性与定量两个方面对实验结果进行分析。本文算法在全部10个实验序列上的平均中心误差为5.3 pixel, 平均覆盖率为77%, 相比于其他算法, 本文算法取得了较高的跟踪精度。本文算法具有更好的鲁棒性, 并且满足更多场景下的跟踪需求。

Aiming at the problem that the subspace representation tracking algorithm cannot deal with the occlusion problem effectively and the sparse representation tracking algorithm cannot meet the real-time requirements of the tracking, this paper proposed a sparse-regularized subspace visual tracking algorithm. The algorithm combined the advantages of subspace representation and sparse representation, improved the processing ability of the occlusion problem, and reduced the computational complexity. Firstly, the algorithm adopted the PCA subspace basis, the subspace mean and the representation residual to represent the target, and used the L2 norm as the regularization of the representation coefficient and the representation residual. Secondly, the algorithm applied an iteration method to compute the coefficients and the residual, then constructed different update templates according to different non-zero ratio of the residual which was preprocessed by opening operator, and employed the incremental principal component analysis method to learn new PCA subspace basis and PCA subspace mean. Through this way, the algorithm enforced the subspace basis and the subspace mean to describe the variation of the target continuously and accurately during tracking progress. Finally, experimental results on qualitative and quantitative aspects analysis showed that average center location error of the proposed algorithm was 5.3 pixels in all 10 experimental sequences, and average overlap rate was 77%. Compared with eight state-of-art algorithms, the proposed algorithm obtains a more precise result and has better robustness and can meet tracking requirements in more situations.