基于稀疏特征直方图约束的鲁棒目标跟踪

Robust object tracking based on sparse feature histogram constraint

为增强目标跟踪的鲁棒性,克服诸如光照、非刚性变化等各种复杂挑战带来的剧烈外观变化,提出一种基于稀疏直方图的通用生成目标跟踪算法。通过滑动窗口将图像分成重叠的子图像块,提取出每一个图像块的局部特征,通过稀疏表示的方法获得对应的稀疏系数向量。为获得特征表达精度,采用增广拉格朗日乘子(ALM)求解最小l1范数问题,目标跟踪中的遮挡问题转换为构建模板的稀疏特征直方图。在粒子滤波框架下,选取候选集中最高相似度指标的图像块作为最优跟踪结果。为适应图标的遮挡带来的形态变化,提出一个简单有效的系数更新机制,产生新的模板直方图。大量基准测试视频定性定量分析结果表明,相比其它跟踪算法,该方法具有更好的跟踪性能。

To enhance the robustness of object tracking, an object tracking algorithm based on sparse histogram with universal generative model was proposed to overcome light, complex background and occlusion challenges. The image was divided into overlapped sub-image blocks using sliding window, and the local features of each image patch were extracted. The corresponding sparse coefficient vectors were obtained by sparse representation. To accurately obtain the expression of the object feature, the augmented Lagrange multiplier （ALM） was adopted to solve the minimum norm problem, where occlusion problem of object tracking was transformed into sparse feature histogram process. In the framework of particle filter, the image patch with the highest similarity measure was selected as the optimal tracking result. To adapt the change of the object occlusion, a simple and effective coefficient update mechanism was proposed to generate a new template histogram. Qualitative and quantitative analysis shows that the proposed algorithm has better tracking performance compared to other tracking algorithms.