基于目标表观和几何建模的物体检测研究及应用
发布时间：2018-08-06 08:58
【摘要】：物体检测是计算机视觉研究领域中一个极富挑战性的课题,是大量高级视觉任务的基础。尽管经历了数十年的研究发展,但在面对实际复杂变化场景时,物体检测的表现仍然存在不足。作为兼具分类和定位任务的复杂系统问题,物体检测始终在模型能力和计算代价的取舍中徘徊前进。针对一般物体检测,本文依据识别对象的几何变化,将目标分为结构化和非结构化目标。结构化物体检测的核心问题是如何表达目标的几何结构信息,如何建模结构化物体的几何变化。针对结构化物体检测,本文假设目标对象的几何变化是透视变换,使用点特征集合表达目标的几何结构,使用S-SVM分类器建模结构化物体检测算法。本文提出预训练和跟踪算法进一步提高结构化物体检测效率。实验结果表明预训练能提高分类器对同一类点特征的辨别力,跟踪算法在不严重损失准确度的情况下能大幅提高检测速度。非结构化物体检测的核心问题是如何表达目标区域信息,如何将候选区域的提取与物体分类统一建模优化。针对非结构化物体检测,本文使用基于特征学习的数据驱动特征来表达目标区域信息,使用Faster R-CNN建模非结构化物体检测算法。本文提出基于多层刺激的候选框融合算法进一步提高非结构化物体检测效率。实验结果表明多层刺激算法能进一步丰富特征抽象能力,该特征使得学习器学习出更加鲁棒的分类规则。综上所述,本文分析了结构化与非结构化的物体检测方法,并提出相应的改进算法,提高了在特定应用场景下的物体检测效率。
[Abstract]:Object detection is a very challenging task in the field of computer vision. It is the basis of a large number of advanced visual tasks. Despite decades of research and development, the performance of object detection still exists in the face of actual complex change scenes. As a complex system problem with both classification and location tasks, physical examination is a problem. For general object detection, this paper divides the object into structured and unstructured object according to the geometric change of the object. The core problem of structured object detection is how to express the geometric structure of the target and how to model the geometric change of the structured object. In view of structured object detection, this paper assumes that the geometric change of the target object is perspective transformation, uses the point feature set to express the geometric structure of the target, and uses the S-SVM classifier to model the structured object detection algorithm. This paper proposes a pre training and tracking algorithm to further improve the physical examination efficiency of the structuration. The experimental results show that the pre training can be proposed. The recognition ability of the high classifier to the same kind of point features, the tracking algorithm can greatly improve the detection speed when the accuracy of the loss is not serious. The core problem of the unstructured object detection is how to express the target area information, how to model the extraction of the candidate region and the object classification, and to detect the unstructured objects. The data driven feature based on feature learning is used to express the target area information, and the Faster R-CNN is used to model the unstructured object detection algorithm. In this paper, a candidate frame fusion algorithm based on multi-layer stimulation is proposed to further improve the detection efficiency of unstructured objects. The experimental results show that the multi-layer stimulation algorithm can further enrich the feature abstraction energy. This feature makes the learner learn more robust classification rules. In summary, this paper analyzes structural and unstructured object detection methods, and proposes a corresponding improvement algorithm to improve the efficiency of object detection in a specific application scene.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
[Abstract]:Object detection is a very challenging task in the field of computer vision. It is the basis of a large number of advanced visual tasks. Despite decades of research and development, the performance of object detection still exists in the face of actual complex change scenes. As a complex system problem with both classification and location tasks, physical examination is a problem. For general object detection, this paper divides the object into structured and unstructured object according to the geometric change of the object. The core problem of structured object detection is how to express the geometric structure of the target and how to model the geometric change of the structured object. In view of structured object detection, this paper assumes that the geometric change of the target object is perspective transformation, uses the point feature set to express the geometric structure of the target, and uses the S-SVM classifier to model the structured object detection algorithm. This paper proposes a pre training and tracking algorithm to further improve the physical examination efficiency of the structuration. The experimental results show that the pre training can be proposed. The recognition ability of the high classifier to the same kind of point features, the tracking algorithm can greatly improve the detection speed when the accuracy of the loss is not serious. The core problem of the unstructured object detection is how to express the target area information, how to model the extraction of the candidate region and the object classification, and to detect the unstructured objects. The data driven feature based on feature learning is used to express the target area information, and the Faster R-CNN is used to model the unstructured object detection algorithm. In this paper, a candidate frame fusion algorithm based on multi-layer stimulation is proposed to further improve the detection efficiency of unstructured objects. The experimental results show that the multi-layer stimulation algorithm can further enrich the feature abstraction energy. This feature makes the learner learn more robust classification rules. In summary, this paper analyzes structural and unstructured object detection methods, and proposes a corresponding improvement algorithm to improve the efficiency of object detection in a specific application scene.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
【相似文献】
相关期刊论文 前10条
1 鲍培明;模糊Petri网在非结构化决策支持中的应用研究[J];计算机工程;2001年12期
2 刘晴;;昆腾创新归档解决方案减轻了主存储上非结构化数据增长的负担[J];计算机与网络;2013年23期
3 缪红;钟华;龙昕;;元器件企业保护非结构化大数据的研究[J];电子元件与材料;2014年06期
4 王晓丽;;基于网络的中学非结构化教育资源建设探析[J];中国教育信息化;2007年08期
5 ToddMatsler;;深耕大数据 助力平安城市智慧转型[J];中国信息界;2013年12期
6 经有国;但斌;张旭梅;郭钢;;基于本体的非结构化客户需求智能解析方法[J];计算机集成制造系统;2010年05期
7 张广泉;非结构化程序流程图及其等价变换[J];重庆师范学院学报(自然科学版);1993年03期
8 王晓波;;非结构化数据采集和检索技术的研究与实现[J];中国内部审计;2014年07期
9 黄远鸣;;一种非结构化数据的多牵度分配存取实现方法[J];科技通报;2014年08期
10 刘威,武家春,廖建新,丁轶;非结构化补充业务数据中心的设计与实现[J];计算机工程与应用;2005年10期
相关会议论文 前10条
1 祝世京;陈s，
本文编号：2167206
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2167206.html
