SIA OpenIR  > 光电信息技术研究室
基于可变形模型的目标跟踪算法研究
其他题名Research on Algorithm of Object Tracking based on Deformable Model
马俊凯1,2
导师罗海波
分类号TP391.41
关键词计算机视觉 目标跟踪 可变形模型 特征选择 尺度自适应
索取号TP391.41/M15/2017
页数107页
学位专业模式识别与智能系统
学位名称博士
2017-11-28
学位授予单位中国科学院沈阳自动化研究所
学位授予地点沈阳
作者部门光电信息技术研究室
摘要本论文首先对现有主流目标跟踪算法做了综述,分析了目标跟踪算法中用到的各个模块在算法中所起的作用。论文针对跟踪中遇到的目标形变、目标尺度变化、目标遮挡、光照变化等问题进行了研究,从图像角度分析了这些因素对目标造成的影响,给出了对应的解决方案。论文的主要贡献如下: 1) 针对目标形变,结合可变形模型与结构化分类器提出了基于可变形模型的目标跟踪算法。目标的可变形模型将目标分解为若干个子块,模型同时衡量了目标局部区域的相似度(局部表观模型)和子块之间空间位置关系(形变模型),使用图模型将两者结合在一起。利用结构化分类器来预测目标中各子块的位置。论文给出了结构化分类器的在线解法,在在线求解算法中给出了一种符合可变形模型形式的核函数表示方法,利用所提出的核函数表示方法可以提高结构化分类器中预测函数的求解效率。实验结果表明该算法可以有效地应对跟踪过程中的目标形变。 2) 针对跟踪中场景环境(光照、背景)变化,提出了一种结合特征选择的目标跟踪算法。在计算机视觉中,不同特征表示的是目标不同方面的信息,如颜色特征关注目标的颜色信息,梯度特征表征目标的边缘细节信息等。在同一幅图像中,不同特征对目标以及背景的表达能力不同;相同的特征在不同图像中对目标和背景的表达能力也不相同。同样是利用目标的局部分块表示,本文提出了一种基于霍夫森林的特征选择跟踪算法。霍夫森林中的每棵决策树与一个特征子集相关联,通过衡量每棵决策树预测能力的大小来确定每个特征子集对目标的表达能力。通过在跟踪过程中逐步舍弃预测能力小的决策树以及补充新的决策树来达到特征选择的目的。实验结果表明了当跟踪视频场景环境变化较大时,该算法仍能准确地跟踪到目标。 3) 针对跟踪中目标发生的尺度变化,提出了一种结合核相关滤波与蒙特卡罗算法的目标尺度自适应跟踪算法。该算法将目标跟踪分为对目标尺度的估计以及对目标位置的预测两个任务。利用核相关滤波算法来预测目标的位置,通过定义目标的尺度变量,给出了利用蒙特卡罗算法迭代估计目标尺度的方法。通过对目标相似度响应图的分析,提出了一种目标模板更新策略,既可以在目标表观模型发生变化的时候更新模板,又能有效防止将错误的跟踪结果作为噪声引入到模板中。实验结果表明了该算法可以正确地对目标的尺度做出预测,在测试数据集上的实验结果表明该算法的跟踪精度优于其他对比算法。
其他摘要This dissertation makes a review on the main target tracking algorithms first and analyzes the role of each module in target tracking algorithms. Aiming at target deformation, scale change, illumination change and so on, this dissertation analyzes the forms of expression in image caused by these problems and comes up with corresponding solutions. The major contributions of the dissertation include: (1) Aiming at target deformation, this dissertation combines deformable model and structured classifier to propose a target tracking algorithm based on deformable model. The deformable model decomposes each target into multi sub-blocks, and at the same time the similarity between the local areas of each target (local appearance model) and the spatial location (deformable modeling) between sub-blocks are also measured and combined by using graph model. The spatial location of each sub-block can be predicted through structured classifier. This dissertation presents an online algorithm based on structured classifier. A kernel function representation according with the form of deformable model is presented in the online algorithm, and it can improve the efficiency of the prediction function in the structured classifier. The experimental results show that this algorithm can cope with the target deformation in tracking. (2) Aiming at scene variation, which contains background cluster and illumination variation, this dissertation comes up with a tracking algorithm based on feature selection. In computer vision, different features represent the properties of the target in different aspects. For example color feature pays attention to the color information of the target, and gradient feature describes the edge information of the target. The same feature can have different expression abilities when refers to the target and the background in the same image. By utilizing the local parts presentation of the target, this dissertation represents a feature selection method in tracking based on Hough Forest. Each decision tree in Hough forest is associated with a feature subset, and the expression ability of each feature subset can be measured by the prediction ability of the corresponding decision tree. In order to choose features, the decision trees with poor prediction ability will be abandoned and new decision trees will be added in the tracking process. The experimental results show that the proposed tracking algorithm can still track the target accurately when the environment of the tracking video scene changes greatly. (3) Aiming at the scale variation of targets, this dissertation proposes a scale adaptive tracking algorithm based on combination of kernel correlation filter and sequence Monte Carlo. This algorithm divides target tracking into target scale prediction and target location prediction. The location of each target can be predicted by using kernel correlation filter algorithm. The scale variant of each target is defined, and the target scale prediction method based on Monte Carlo is presented. After analyzing the graphs of target similarity, this dissertation presents a target model update strategy, which can update model when there are changes in target appearance and avoid introducing incorrect tracking results into models as noises. The experimental results show that this algorithm can predict the target scale correctly. Comparison results show that this algorithm outperform other state-of-art methods.
语种中文
产权排序1
文献类型学位论文
条目标识符http://ir.sia.cn/handle/173321/21273
专题光电信息技术研究室
作者单位1.中国科学院沈阳自动化研究所
2.中国科学院大学
推荐引用方式
GB/T 7714
马俊凯. 基于可变形模型的目标跟踪算法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2017.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
基于可变形模型的目标跟踪算法研究.pdf(29129KB)学位论文 开放获取CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[马俊凯]的文章
百度学术
百度学术中相似的文章
[马俊凯]的文章
必应学术
必应学术中相似的文章
[马俊凯]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。