基于角点的稀疏子空间粒子滤波跟踪算法

PARTICLE FILTER TRACKING ALGORITHM BASED ON CORNER FOR SPARSE SUBSPACE

大多数现有的稀疏表示目标跟踪方法计算复杂度高,对琐碎模板采用直接将整个图像块按顺序划分成若干个小块放入模板中,引入了背景信息,容易跟踪漂移,甚至丢失目标的情况。针对这种情况,提出基于角点的稀疏子空间粒子滤波跟踪方法。通过Shi-Tomasi角点的方法提取特征点,选择其中特征点最多的图像块建立琐碎模板,对相似度高的仅采用基向量对候选目标进行误差运算,增量更新基向量子空间,否则将正交基向量结合琐碎模板组成字典并加入从高到低的权值进行误差计算,同时更新琐碎模板。与当前多个目标跟踪算法相比,定性和定量实验结果均表明,该算法在目标发生遮挡、旋转、尺度变化、快速运动、光照变化等复杂情况下跟踪效果更好,实用性更强。

Most of the existing object tracking methods based on sparse representation have high computational complexity. For trivial templates, it is easy to track drift and even lose the target if the whole block is directly divided into several blocks into a template, which will introduce background information. In this case, a sparse subspace particle filter tracking method based on corner points is proposed. Extracting feature points through the Shi-Tomasi corner point method, then choosing the image block with most feature points to establish trivial template. For high similarity, only base vectors are used to perform error operations on the candidate targets, and the base vectors subspace is updated incrementally. Otherwise, the orthogonal base vector is combined with a trivial template to make up the dictionary and add the weight, which is from high to low, to calculate the error and update the trivial template at the same time. The qualitative and quantitative experiment both show, compared with the current multiple target tracking algorithms, the proposed algorithm is more precise and practical in complicated situations, such as occlusion, rotation, scale change, fast motion and illumination change.