基于改进协作目标外观模型的在线视觉跟踪

发布时间：2018-06-01 06:26

本文选题：在线视觉跟踪 + 协作外观模型　；参考：《电子学报》2017年02期

【摘要】：在不受限制的复杂环境中在线跟踪任意类型的感兴趣目标仍是一项极具挑战的难题.本文在无模型跟踪框架基础上提出一种基于改进协作目标外观模型的在线视觉跟踪方法,解决了大多数协作模型类跟踪算法在学习阶段无法有效选择正、负样本的问题.该方法根据人类视觉感知准则将目标边缘信息视为最具区分度的目标特征,提出边缘判别模型并结合动态模型和检测模块建立二级似然匹配空间,为生成模型的似然匹配去除了背景干扰;采用分块策略建立目标生成模型,为模型引入空间结构信息;利用Mean-Shift计算各子块的最终位置和匹配系数,并根据子块匹配系数为遮挡处理和模型更新提供依据.在公开视频序列上同几种流行视觉跟踪算法的对比实验结果证明了本文算法的有效性和优越性.
[Abstract]:Tracking arbitrary types of objects of interest online in an unrestricted and complex environment remains a challenging challenge. In this paper, an online visual tracking method based on improved collaborative object appearance model is proposed on the basis of model-free tracking framework, which solves the problem that most collaborative model class tracking algorithms can not effectively select positive and negative samples in the learning phase. According to the human visual perception criterion, the target edge information is regarded as the most discriminative target feature, and the edge discriminant model is proposed and combined with the dynamic model and the detection module to establish the two-level likelihood matching space. The background interference is removed for the likelihood matching of the generated model, the target generation model is established by block strategy, the spatial structure information is introduced into the model, and the final location and matching coefficient of each sub-block are calculated by Mean-Shift. According to the subblock matching coefficient, it provides the basis for occlusion processing and model updating. The effectiveness and superiority of this algorithm are proved by comparing with several popular visual tracking algorithms on open video sequences.
【作者单位】：信息工程大学信息系统工程学院;南京理工大学自动化学院;
【基金】：国家科技重大专项(No.2014ZX03006003)
【分类号】：TP391.41

【相似文献】