摘要:
目的：准确性和鲁棒性是视觉跟踪的主要挑战，尤其是在遮挡和目标形变的情况下。对目标进行分块跟踪能够保持目标的空间信息，基于此，本文提出一种融合多核的目标分块跟踪方法。方法：算法采用垂直投影法将目标分成若干合适的子块，针对相应子块，选择目标区域内的多个不同位置构建多个核函数加权直方图，以Bhattacharya系数表征目标模版和候选目标模版之间的相似性度量，并进行均值迭代确定目标最终位置。在跟踪过程中，利用分量的反投影运算区分目标发生形变或遮挡，并对目标模板和子块权重进行实时更新。结果：对多个视频序列进行了实验测试，实验表明，该方法几乎不受光照影响，并且在目标大面积遮挡的情况下仍能实现很好的跟踪。结论：通过将分块和多核融合结合起来进行跟踪，不仅对光照不敏感，而且在处理目标大面积遮挡方面鲁棒性很强，利于后续论题的研究。
Abstract: Objective: Accuracy and robustness are the main challenges of the visual tracking, especially in the case of occlusion and target deformation. Target tracking based on fragments will be able to keep the spatial information of the target, and based on this, a fragments-based tracking algorithm with multiple kernels fusion is proposed in this paper. Method: Vertical projection method is used to get proper fragments in the algorithm, and for each corresponding fragment, it selects a plurality of different locations within the target area to build several kernel function weighted histograms, taking the Bhattacharyya coefficient as the similarity measurement between the target template and the candidate template, and making use of the mean shift iteration to determine the final po-sition of the target. In the process of tracking, it takes advantage of the back-projection of the components to distinguish deformation or occlusion, and makes a real-time updates for the target template and fragments weight. Result: According to the results, which are obtained from several testing of video sequence, the method is almost not influenced by illumination, and can still achieve good tracking even in a large area of occlusion. Conclusion: The proposed algorithm, by combining fragments and multiple kernels to tracking, is not only insensitive to illumination, but also has a good performance in dealing with a large area of occlusion, which is beneficial to the research of next stage.