懒交互模式下散乱不规则分块引导的目标跟踪

Visual tracking based on lazy interaction and irregular patches

针对目标运动过程中的复杂背景和光照变化等挑战展开研究,提出基于懒交互和散乱不规则分块的目标跟踪方法,以提高目标跟踪的准确度和成功率。首先,采用简单易操作的懒交互方式将目标分成多个散乱分布的不规则分块,并用核相关滤波对每个分块进行初始化建模;然后,基于核相关滤波对每个分块进行跟踪。同时,为了适应目标和环境的不断变化,先对分块模型进行简单更新,当简单更新无法满足目标变换时,对相关分块进行懒交互式重采样,以构建更准确的目标分块模型。最后,根据所有分块在新一帧的位置,根据霍夫投票确定目标位置。主要创新为:通过懒交互的方式采样能够保证分块的有效性;通过对分块进行散乱不规则的采集能够对目标特征进行有效典型的描述。针对Visual tracker benchmark的27个视频进行测试用于评估不同跟踪方法,实验结果表明,本方法处理光照变化、旋转、复杂背景时能够得到更精准的跟踪结果。

A novel visual tracking method based on lazy interaction and irregular patches is proposed, to improving the accuracy and success rate of target tracking, especially on the challenges of background clutters and illumination variation. First, it divides the target into a lot of irregular target patches by using a simple and lazy interactive process, and each target patches is initialized by the kernel correlation filtering. Then, each target patches is tracked based on the kernel correlation filtering. At the same time, the proposed method updates the patches model in a simple way to satisfy the constant changes of the target and the environment. When a simple update fails to deal with the target transformation, the lazy interactive process resamples the related patches and builds a more accurate target patch model. Finally, according to the position of all the target patches on the new frame, the target position is computed by the hough-voting scheme. The main innovation is that sampling the target patches through lazy interaction ensures the effectiveness of describing target patches. And by sampling the irregular patches, it effectively describes the target characteristics. Using the 27 videos of visual tracking benchmark, the quantitative evaluation of different trackers is achieved. Many experiments demonstrate that the proposed method perform is better in dealing with target illumination variation, rotation, background clutters and so on.